Last week Brad and Will got to attend a virtual deep dive on Nvidia's new 30-series GPUs, and in this ep we attempt to break down all the new ray-tracing optimizations, DirectStorage capability, whole-system latency reduction, push-pull cooler designs, V-shaped PCBs, and more. Be warned, it's a lot! Support the Pod! Contribute to the Tech Pod Patreon and get access to our booming Discord, your name in the credits, and other great benefits! You can support the show at: https://patreon.com/techpod
Last week Brad and Will got to attend a virtual deep dive on Nvidia's new 30-series GPUs, and in this ep we attempt to break down all the new ray-tracing optimizations, DirectStorage capability, whole-system latency reduction, push-pull cooler designs, V-shaped PCBs, and more. Be warned, it's a lot!
Support the Pod! Contribute to the Tech Pod Patreon and get access to our booming Discord, your name in the credits, and other great benefits! You can support the show at: https://patreon.com/techpod
Will: [00:00:00] I'm not sure about these wheels, Brad,
Brad: [00:00:01] Why, but I was told that the glide smoothly, that they are silent, that they are the best thing since sliced bread.
Will: [00:00:09] there's a little bit more startup inertia than I'd like, like there's a little like, well, they rotate. Maybe I got, maybe the ones I got were too cheap, but I don't, my old wheels were pretty good. They just
Brad: [00:00:20] Your speed off the line. Your zero to 60 is suffering. Is that what you're saying?
Will: [00:00:23] my straight line speed is real good. Okay. It's like it's.
Are
Brad: [00:00:26] you doing street races in your office chair. Like what is the problem here.
Will: [00:00:29] I haven't put enough points into spin. Okay. I've been, I can't think about anything else. I played eight hours of Tony Hawk yesterday. I can't think of anything, but Tony Hawk, I didn't play eight hours of Tony Hawk, but I did. I did get it working on Moonlight so I can play Epic store games in my living room, in my bedroom now. Yeah. I like having a breaks on the chair wheels. I've never had that before.
Brad: [00:00:50] Wait,
Will: [00:00:51] Yeah I bought I bought rollers Would rollerblade wheels with brakes
Brad: [00:00:54] With breaks
Will: [00:00:55] Yeah With brakes
Brad: [00:00:56] How do you trigger the breaks Is it on the
There's
Will: [00:00:58] a little lever You push it's like it's like a foot brake on a rolling cart
Brad: [00:01:01] Wow let's do to lock your chair into place
Will: [00:01:03] I can lock my chair And I figured in case I got really into racing Sims or maybe like get rudders for my HOTAS situation Yeah I'm you know I'm good to go
Brad: [00:01:12] So what would that let me honest appraisal
Will: [00:01:15] So there's no side to side wiggle Once you once you roll into place which is nice
Brad: [00:01:20] cause I'm doing it right now in this chair and it feels terrible
Will: [00:01:23] Yeah But when you when you when you push back there's a little bit of a moment while the wheels flip back around to be the aims of the right
Brad: [00:01:29] okay Sure
Will: [00:01:31] and it adds a little bit of drag that I'm not I was expecting this to be butter smooth all the way through
Brad: [00:01:36] Well so it actually now that you mentioned it like the normal caster on a traditional office chair is it's two it's actually two wheels right That can turn independently Huh
Will: [00:01:44] And it turns out the Steelcase wheels When I was looking at reviews of these things they're like Well my Steelcase wheels Cause when you buy a gesture a nice Steelcase chair they say what kind of floor do you have Do you have a hardwood floor do you have carpet floor And they give you the right casters Um it's it's the Castros were pretty nice So it may just be that I had real good wheels already and didn't know it And I replaced them with slightly less Good but more maybe more functional
Brad: [00:02:12] What if what if one were still sitting on casters from let's say office max circuit 2003
Will: [00:02:19] Look Brad do you need to just put the chair and the old monitors out on the curb with a note that says free to a good home and then walk away
Brad: [00:02:27] That was the plan but I have a taker on those monitors Somebody has signed up
Will: [00:02:32] is it A fan
Brad: [00:02:33] no it's someone from work I mentioned it in the newsletter that I wrote for the giant bomb newsletter I wrote two three two weeks ago three weeks ago I just happened to mention like Oh we're getting all this new gear from the budget stuff that we've got a surplus and we need to gear up at home So what am I going to do with these 15 year old monitors and someone slacked me is that's the weird thing about working at a giant company with a company wide Slack account is like anybody in the company can get to you very quickly
Will: [00:02:59] Was this Stephen Colbert
Brad: [00:03:01] It was not Stephen Colbert
Will: [00:03:02] Are you sure
Brad: [00:03:04] I don't think Stephen Colbert is in that database I might be wrong I could be wrong I could pull I shouldn't do this I could pull it up right now
Will: [00:03:12] Was it the numbers lady
Brad: [00:03:14] That's a FOX thing right
Oh well someone from another department who I had not met before but uh should I should I really shouldn't do this I'm clicking on the direct message
Will: [00:03:22] button right now
No this is a bad idea
Brad: [00:03:24] now Steven is it a pH or a V
Will: [00:03:28] Look it's impossible for me to know the answer to this question
Brad: [00:03:31] I have an answer I don't know if I should say it
Will: [00:03:33] Yeah you can Slack Stephen Colbert
Brad: [00:03:37] There is there is an entry in here I can't say
Will: [00:03:39] should ask him to come on the show We can talk about the fall of America It'd be great
Brad: [00:03:44] I highly doubt that those messages go straight to him but
Will: [00:03:46] I bet it does I bet you could I bet you can make his phone ring right now
Brad: [00:03:49] whoever is managing that account it says they're online right now
Will: [00:03:53] Look there Stephen called is the most always online person in the world He knows everything
Ain't
Brad: [00:03:59] that the truth
Will: [00:04:00] I love that guy Welcome to Brad and will made a techpod I'm Will
Brad: [00:04:35] I'm Brad
I'm
Will: [00:04:37] glad we were able to carry forward the long tradition of finding out which famous CBS people you're able to access for your interoffice communication tools
Brad: [00:04:45] I don't know what it says about me but I mean like I you know like nothing against Stephen Colbert but I get more excited looking up like say 60 minutes personalities
Will: [00:04:52] Yeah Well I mean look
Brad: [00:04:54] hardcore news types
Will: [00:04:56] if if uh
Brad: [00:04:57] Jeff Glor is not online right now
Will: [00:05:00] Would you do me a favor and let Julie Chen know that the current season of big brother is a real time capsule into our COVID lifestyle and I'm quite enjoying it
Brad: [00:05:07] Sheet she's also in here
Will: [00:05:09] Of course she is
Brad: [00:05:10] we have to stop talking about
Will: [00:05:11] this
Yeah This is a bad idea
Brad: [00:05:12] I will not I will never message any of these people just for the record for anyone anyone in management listening
Will: [00:05:17] this is a bit this is very much a bit um what are we going to talk Oh first off we have news Something magical has happened
Brad: [00:05:25] what happened
Will: [00:05:26] This week we hit 1000 backers of the tech pod Patrion We don't usually talk about this at the beginning to annoy the people who don't aren't paid to patrons which is fine We love you guys too
Brad: [00:05:35] Yeah but there's housekeeping associated this one
Will: [00:05:38] We had a challenge We issued a challenge that if we hit a thousand backers of the tech pod Patrion uh then we would uh Brad who has never read Hitchhiker's guide to the galaxy and I who have read Hitchhiker's guide to the galaxy many many times
Brad: [00:05:50] How many times would you estimate
Will: [00:05:52] Oh dude Alright It's I read it first when I was 12 years old And for a while I was reading it like you know every few years
Brad: [00:06:02] Okay I get
Will: [00:06:03] it
Um I've listened to the radio plays I've watched the eighties TV show I've watched the movie I've listened to the abridged audio version the radio dramas not the not the radio plays but the abridged audio book versions the dramatize ones I've listened to the unabridged audio book versions I think that's all the main forms
Brad: [00:06:22] You've consumed every possible permutation You know it would be a fun experiment which I am not going to do to be clear but what if I were to consume one of those other versions Like what if I just watched the recent movie and then showed up ready to talk about the book
Will: [00:06:35] Oh Brad
Brad: [00:06:36] don't do that
Will: [00:06:38] So I like the recent
Brad: [00:06:40] movie
how well I could keep up with that discussion and I think it would be kind of a fascinating exercise but I but I also I fully acknowledge that I would be cheating myself by doing that
Will: [00:06:50] Yeah I th I think start with starting I thought about this a lot I think starting with the book is good The radio
Brad: [00:06:55] player are really good
No no I
Will: [00:06:56] a really good
Brad: [00:06:57] intend to read the book
Will: [00:06:58] The movie the movie is one of those movies that like if you weren't familiar with the source material is a pretty fucking good movie but when you compare it against like the radio plays and even like the like it's like Martin Freeman and Zooey Deschanel and most Def are great but
Brad: [00:07:13] not a bad so it's not a bad cast I didn't I didn't
Will: [00:07:15] and uh what's his name for moon Um
Brad: [00:07:18] Sam Rockwell
Will: [00:07:19] Sam Rockwell our greatest living
Brad: [00:07:20] actor
Sam Sam Rockwell is incredible
Will: [00:07:22] Yeah he's very good It's like it is it is a well done thing but it isn't it misses a little bit of what makes hitchhikers special Cause it's a little you know it's it's it's it's you know imagine if they made an Americanized version of Shaun of the dead
Brad: [00:07:37] Okay sure
Will: [00:07:38] right That's kind of what you end up
Brad: [00:07:40] Okay Enough said I I'm good Uh I mean yeah something something this hallowed and Seminole I would never experience in any form other than the original
Will: [00:07:50] So you found your omnibus I have my digital version We're going to start reading Now We're going to kind of shoot for maybe first episode of October is this so this is your this is your warning This is this is not a patron exclusive anything This is for everybody The patrons have unlocked the achievement for all of us Um we'll do it as the first episode of
Brad: [00:08:08] October
what is that episode Cause it's just a book club
Will: [00:08:12] I don't know
Brad: [00:08:13] Oh we'll figure it
Will: [00:08:13] it will tell you we got a month We'll figure it out
Brad: [00:08:15] Okay
Will: [00:08:16] Uh but this week the business at hand is a little tried to come up with segway here and I don't have it
Brad: [00:08:25] this is a dense episode dude New graphics cards
Will: [00:08:28] Yeah New video cards and immediate announced
Brad: [00:08:30] that's the most wonderful time of the year
Will: [00:08:33] We've been waiting a long time
It's
Brad: [00:08:34] that that magical moment in a young man's life when
We get
Will: [00:08:38] invited into Jensen Wong's kitchen and we get to see where he bakes up the video cards
Brad: [00:08:44] I'm glad you mentioned that Uh they acknowledged the spatulas
Will: [00:08:48] Oh yeah Hold on First up a disclosure uh Nvidia paid me to talk over their stream earlier this week So uh you know they asked for honesty uh natural feedback which is what I provide when people ask me to do that sort of thing Uh and but I have to disclose that because you know it's the law so yeah
Brad: [00:09:09] I suddenly feel this like thin layer of film on top of me I don't know where where's this coming from It's like this it's like a
Will: [00:09:15] Is it food Is it like a residue from eating
Brad: [00:09:17] a no it's this is vaguely slimy No I'm just I am just joshing You
Will: [00:09:23] poor me
Brad: [00:09:24] Uh to be clear I was not part of that deal
Will: [00:09:27] No Brad was not part of that deal
Brad: [00:09:28] I they did not they did not pay me but uh but but they did invite me along with you in into I don't know what they would typically call this like kind of their their their deep dive
Will: [00:09:38] They call them editors days is what they usually call them
Brad: [00:09:40] Okay Okay So that I mean I'm very familiar with that uh concept from the games press uh the biggest differentiator here is how insanely technical this entire thing was
Will: [00:09:51] So these things started years and years ago Um when I was at maximum PC and reviewed video cards I went to them every pretty much every year for every new card Um they in the old days we would get the new cards and then we would look at what they said the cards were capable of And we'd do a bunch of math and run some weird benchmarks and try to figure out how the cards worked inside And then eventually they were just like just stop doing that We're just going to tell you how stuff works.
Brad: [00:10:14] Stop getting it wrong We're just going to give you the information straight
Will: [00:10:17] And they give they give I would describe like it's so It's one of those things where like the people from digital foundries are there the people who do like goober 3d and all those guys are there And it's it's like the people who the range of people who are there are from like people who have C CS degrees in graphics rendering and stuff like that and have written SIGGRAPH papers all the way down to like me and you Brad And You get an incredible amount of detail So we're going to take like six hours of briefings and we're going to distill that down into a 45 minute podcasts that normal humans can understand
Brad: [00:10:51] Oh it was how long was hold up let me check here Two to four
Will: [00:10:56] We Started at eight o'clock and finished it two
Brad: [00:10:59] we wrote nine pages of notes off of those presentations single space Um but uh yeah the very first slide of the first presentation was a spatula joke I just wanted to say so they are somewhat self aware about this stuff.
Will: [00:11:13] I went back and looked and you could see the card there behind the spatulas the whole time Yeah Also the spatulas were in the original ampere rollout too when they did like the computational ampere here in the spring
Brad: [00:11:25] You know it's like I got nothing but respect for a man who loves his spatulas
Will: [00:11:29] I'm my next next time I go down there I'm going to make a I got a I'm going to cast a custom spatula and take it as a gift cause uh
Brad: [00:11:36] That's pretty good I you know we could we could use some more specialists around here Quite frankly You can never you can never have enough I like to make eggs Uh but yeah I just wanted to set the stage a little bit I I saved a handful of the slides from this thing The first one I had like five minutes into it They were talking about the like digging deep deep into like the ray-tracing workload pipeline And like I've got slides here talking about bounding volume hierarchies and linear equations and you know
Will: [00:12:05] Well so so it's one of the things like when when when it's easy for them to say at at at a standup presentation Hey our our ray tracing core is are two times faster than they were in turing in the 2000 series
Brad: [00:12:17] yes But this was a super useful cause like I got a much better sense of why they're that much faster
Will: [00:12:22] Like they sit down and explicitly say look to make the the Ray tracing course faster We they they figured out where the bottlenecks were from watching you know data on dozens of games that have come out and thousands and thousands of people using it to see where the bottlenecks are And then they designed this new stuff to bypass those bottlenecks So For example things like the there are units that can do the math that calculates whether triangles or bounding box are intersected by rays. Like that is a hardware accelerated bit now And all of these things can run concurrently So previously they had to run sequentially and now they run concurrently and that turns out and it makes things faster
Brad: [00:12:59] Yeah I think I think the kind of jist we got from that big hardware presentation and I think then one of us wrote in these notes somewhere is that I think it was you that wrote that like this gave the impression that you know the the previous generation of cards was the first rollout for hardware based ray tracing at all And then this gives the impression that they looked at the how they accomplish that and saw a bunch of very glaring bottlenecks that they could easily solve
Will: [00:13:22] I mean that's kind of the story of graphics in general over the last 20 years but but it's definitely that that was the feeling I got
Brad: [00:13:29] This was very pronounced you know because they were like diverting entire functions off to like kind of co-processing hardware or you know.
Will: [00:13:36] Like the units in the in the cores and
Brad: [00:13:38] that's that's what I mean
Will: [00:13:39] so the difference between Um uh uh Pascal touring and ampere So Pascal Pascal is the 1000 series cards Uh touring is the 2000 series of cards and ampere is the 3000 series cards Um is that like for example on a 1080 when you do Ray tracing if you force Minecraft to do Ray tracing on a 1080 It can do it but it runs all of the math on the shader processors If you do it on the on the ampere and the Turing cards it uses the the RT accelerators which are sections of each Um each of the cores on the on the GPU and there are there are thousands of these things like there's there's they're literal buttloads Um and so it's something like uh accelerating linear equations which is something that that hasn't necessarily needed to be accelerated in hardware on a GPU before and now they can run a whole lot of those concurrently because they use those to do the bounding calculations or the the the the collision calculations for whether the rays from light sources impact Specific triangles So
Brad: [00:14:43] this this was really illuminating for a layman such as myself I mean I always knew this intuitively but to actually see it all laid out that like every single thing going on in the construction of a frame is just a math problem You know like when you really break it down like what is this hardware doing It's just as much as spoon we've talked about this
Will: [00:14:58] an insane amount of specialized math
Brad: [00:14:59] just a bunch of specialized hardware to solve very specific types of math problems And you know you solve how however many what millions of those per frame or something
Will: [00:15:09] Well I mean they were talking about 8K gaming which is like 33 million pixels a frame or something like that It's bonkers Um so yeah let's talk about the cards there There are three cards that they announced the 3070 the 3080 and the 3090 And I'm realizing now that I didn't write prices down for any of these which
Brad: [00:15:26] It's uh I I've got it in the old noggin Uh 3070 is four 99
Will: [00:15:33] Okay So 500 bucks is the starting point
Brad: [00:15:34] Yes 3080 is six 99
Will: [00:15:37] So 200 bucks up
Brad: [00:15:38] Yes And then the 3090 is 1499 So
Will: [00:15:43] uh eight 700 bucks up.
Brad: [00:15:44] More
Will: [00:15:45] Say 800 bucks
Brad: [00:15:46] Yes more than double the next step down Like the 3090 is the new Titan I think they more or less came out and said that
Will: [00:15:51] said that yeah they explicitly said
Brad: [00:15:53] They explicitly said this is the card for crazy I shouldn't say crazy people They're like there are definitely legitimate applications for it but it is like You know you're you're not by more than doubling the price of a 3080 You are not more than doubling the performance or anything close to that
Will: [00:16:08] Well it depends on what you're doing Cause if you're doing stuff that's like math like like they rolled out the Titan last, I mean they've always had the Titan for a fairly long time now but they assumed up until this point it was for people with really specific workloads like computational math people or AI people doing AI models on their on their workstations and stuff like that And it turns out a lot of people bought Titans last year for more traditional workloads Um and as the benefits of SLI have waned Uh progressively more each generation to the point that like nobody should fucking run SLI now unless you have a explicit need for it Um they the benefit of having a big chunker of a card so you can play games at 4k this generation or eight K on the next generation um makes a little bit more sense Um it's it is I the the thing that is telling about this generation is the first time in since like in like 10 years maybe more than that that I remember the low end card from the next generation being faster than the high end card from the previous generation
Brad: [00:17:10] that is that in fact the case 3070
Will: [00:17:12] so there are two big takeaways If you don't care about any of this other stuff the one is that they acknowledged that on the last generation they focused on bringing RT Ray tracing and getting that working But you didn't see traditional game performance boost
Brad: [00:17:28] Yeah I
Will: [00:17:29] things got better looking but they didn't run faster
Brad: [00:17:31] I was shocked at how candid uh Jensen Wong was in that presentation to stream
Will: [00:17:36] They're very like they're they're pragmatic
Brad: [00:17:39] That line toward the end where he basically was just like all my Pascal friends out there it's safe to upgrade now Like you're basically tacitly acknowledging that like Hey maybe the previous generation was not really worth it for people who just wanted higher frame rates
Will: [00:17:51] Well it's kind of the thing we said like Hey this one is great If you want Ray tracing but you know the benchmarks the benchmarks the benchmark improvements over previous generation cards were a little underwhelming on the 2080
Brad: [00:18:05] Yes this is this is the I mean I don't know if you can quite apply the old Tick-Talk model to this thing but you know last time was a feature upgrade This is a performance upgrade
Will: So these would both be ticks So there's the super is would be the talk on on the tick talk model metaphor
Brad: [00:18:19] Well maybe that acknowledging doesn't fit Exactly But like what point is like I mean there are new features to this but in terms of the basic graphical capabilities it's kind of doing the same stuff faster
Will: [00:18:27] They're kind of all yeah I mean yes it is definitely It is definitely Here's what we learned from the first go on this and let's make it better this go around.
Brad: [00:18:34] Um yeah and again seeing seeing how all the pipeline has been optimized and you know new acceleration added and X Y and Z really it just gave a lot of good background to why the performance increases so profound here Uh it it made it very clear like Hey this is not smoke and mirrors This is exactly why these are so much faster They are they're doing all these math operations in an incredibly more efficient and more parallel way
Will: [00:18:58] Well and they're also starting to think about other places in the GPU in the in the game pipeline that the GPU can help speed pain points and remove bottlenecks Um which you know like I said graphics is all about removing by it's like a video a video card in the basic sense is just about removing a bottleneck That the CPU is bad at and moving it onto a dedicated coprocessor anyway Okay So 3070 is 1.7 These are all about 1.7 Gigahertz is tiny Cox speed differences I'm not gonna get into that The 3070 has eight gigs of GDD are six Ram and 58 88 5,888 CUDA cores So CUDA cores includes all of the specialized like math courses So that that I believe is going to include um Uh RT tensor and um uh shader cores um
Brad: [00:19:49] Tensor cores are the AI accelerators Is that right
Will: [00:19:50] tensor cores are machine learning accelerators So they're um they're uh they're basically mimicking like neural nets So your your training um a matrix of possible outcomes and the ones that are weighted heavily get used in the ones that are weighted lightly get culled and then you move on to the next thing
Brad: [00:20:11] They have they mentioned they have acceleration or optimization for that process too now
Will: [00:20:15] Yeah So the optimization for that I don't I'm a hazy on how all the machine learning stuff works Cause it's it's I mean in one on it's kind of just statistics but it seems like there's more to it than that Cause it gets really complicated really fast
Brad: [00:20:28] Stuff seems like more than anything like the domain of pure academia to me or like that
Will: [00:20:33] Well kind of but it's showing up in games and
Brad: [00:20:35] Oh I know it is now but what I
Will: [00:20:36] using it everywhere
Brad: [00:20:37]It a very obvious pipeline straight from the research world into practice
Will: [00:20:41] It's it's one of those things that like somebody somebody in in who knows more about this is going to tell me I'm wrong on this probably but it feels like one of those things that like you remember reading science fiction stories where the people stopped understanding how the computer programming works because they started having the computers program themselves It feels like the first step on that patch
Brad: [00:21:00] Oh absolutely
Will: [00:21:00] I know it's not but like it feels like it is because we get these weird algorithms that do what we want We don't really understand exactly how they work Um Okay So uh 3070 has 55,888 CUDA cores eight gigabytes of normal GDDR6 Ram The 3080 has 8,700 CUDA cores with 10 gigahertz gigabytes of GDDR6 X Ram which is new Um and the 3090 has 10,496 Cooter cores and 24 gigs of GDDR6 X Ram
Brad: [00:21:35] So that's actually the biggest bump on the 3090 is the amount of Ram which is way more than double the 3080 but like the the uh the Cuda cores the difference in crude cores is not gigantic
Will: [00:21:45] well it's a 20 it's a 25% increase which is substantial
Brad: [00:21:49] It's not it's not for sure
Will: [00:21:49] but it's almost 25%
Brad: [00:21:51] nothing but again like like I said when you're talking about a card that costs more than double what the other card does
Will: [00:21:56] Yeah I mean the benefit is the Ram They were talking about the having multiple like giant ML workloads running at the same time and stuff like that The secret of a lot of the machine learning stuff is that if it runs out of memory it goes infinitely slower is my understanding But I don't I don't know about that So uh all these all these chips are on Samsung's eight nanometer process and um the the the so the GDDR six X is interesting Cause they're I think they're doing more memory ops per clock cycle Would that then with traditional typical GDD are six Um
Brad: [00:22:29] specifically on the X variant is what you're talking about Okay
Will: [00:22:33] is is and it seems like from what, I can't remember it was Tony Tommasi and the deep dive or Jensen in the in the main thing it seems like we're maybe reaching a point where we're going to see diminishing returns on memory bandwidth which is going to have some interesting complications um
Brad: [00:22:50] You have written here four signal levels per cycle in each cell versus two in the previous type like
Will: [00:22:55] is that that means there
Brad: [00:22:56] Is that the would surely that's not the like literal electrical signal
Will: [00:23:01] That's the literal electrical signal
Brad: [00:23:03] How what is is
Will: [00:23:05] They’re sampling stuff on this so that they get accurate results It seems it seems like there's some real there's some real weird logic happening on this.
Brad: [00:23:11] was I was gonna I was gonna say like on and off are very discreet States that you can read with pretty high confidence Right But if you start getting into weird variations of on that does that seems I mean obviously they've made it work but that sounds magical and risky to me Okay
Will: [00:23:26] I don't understand how that stuff works Brad it's magic and risky
Brad: [00:23:30] Okay there’s going to be several examples of that throughout this stuff
Will: [00:23:33] Like the danger of going to these editors days is that you can generate 10 pages of notes in six hours and then immediately go back and turn each page of those notes into another 10 or 20 pages of notes If you start doing a lot of reading Um so so the the big the big performance per watt stuff the interesting data here is they're saying and always these are Nvidia provided numbers so skepticism but but Also the people have the cards now and can do the benchmarks themselves So you find out pretty quickly if they're cooking numbers um it's a 1.9 X performance per watt improvement under RT workload So that means if you are playing a game like control you're going to see two times the performance for each watt of energy consumed by the card Which means what this means is you'll get a playable frame rate without having to have the blower going full speed Um at the resolutions most people play games at a 1080P, 1440 and probably even 4k Right
Brad: [00:24:31] Yeah I was a I was surprised in the graph They showed that the it seems like touring the previous architecture uh more or less plateaus at about 240 Watts according to their graph
Will: [00:24:42] I think that's right Yeah
Brad: [00:24:44] essentially that was the ceiling on that on that hardware And I assume that's due to all those inefficiencies in the pipeline that they have is to some degree
Will: [00:24:51] I'm also probably cooler design Like there's there's the you know you're limited by what a lot of this stuff that you know you have to decide how hot you want the card to run and how much heat you want to generate Cause that's going to determine the life span of the card Um but yeah and and what the what the cooler can actually blow out And the 28 having used it for a while the 2080 cooler just sucks
Brad: [00:25:12] Okay
Will: [00:25:13] It's not great.
Brad: [00:25:14] I've never used one but that's good to know So I mean but the upshot for the label the lay person is that you could not keep pumping more energy into the 2080 series cards and get better performance Like you hit a pretty low ceiling on that Whereas these these cards continue to scale with more power more and more electricity being pumped into them significantly higher than the than the old ones did
Will: [00:25:34] You have, my actually I think about it more with the touring cards probably where they're capping out because there's some bottleneck that you can't clock past.
Brad: [00:25:39] That's what I meant
Will: [00:25:40] you can't be like pump more more speed Doesn't make it faster at that point
Brad: [00:25:46] Like not to get too much into the math but when they were talking about you know the bottlenecks in there how they cast on sample rays for all the ray-tracing operations they were talking about from what I what I got from it sounded like they were doing a lot of that in kind of a linear or sequential way that was not very efficient And now they have shunted all of that off to the side so that it happens
Will: [00:26:03] Well it can happen independent of the shader pipeline which is the key thing
Brad: [00:26:06] more in parallel And like those are the types of bottlenecks that you wouldn't be able to just brute force your way past Right
Will: [00:26:11] That's exactly it And and that the idea of decoupling a lot of the different parts of the pipeline is is an interesting I wouldn't have thought that like traditionally you think about the graphics pipeline of being like triangle set up which is where it makes the mesh that gets rendered And then it applies some textures and there's a lighting pass and all that stuff And and with the way this is working is I wouldn't have thought as possible but they said they were able to figure it out They're there It's funny when you get down to this stuff there's also a lot of weird temporal like you've heard about temporal AA which basically looks at frames ahead and behind uh of of your device and smooth based on what's coming Not necessarily what's on the screen at any given moment And I feel like they're doing a lot of that stuff for for some of these features like motion blur and stuff now too Um uh so uh would you say scales better with TDP The upshot is well a they designed a new cooler which you probably have already seen I I'm incredibly curious to see how this thing performs Uh they're saying at uh at ampure is running at 78 Celsius at 30 decibels for the for the um for the cooler
Brad: [00:27:23] is that at a given at a given performance point
Will: [00:27:26] Well that's that's of when it hits 70 AC for the card That's the temperature That's the loudness of the fan
Brad: [00:27:33] Oh sorry I'm getting ahead of you Cause it's cause they also have touring comparisons here
Will: [00:27:37] Yeah So turing was running at three degrees Warmer was running This is this is for the same performance profile
Brad: [00:27:45] I was wondering what yeah that sounds trying to figure out what what what the axis was here
Will: [00:27:48] This is essentially a half power ampure equivalent to a full speed Turing And the Turing is 32 decibels at 81 C at the same same like a power input
Brad: [00:27:59] Yeah Yeah So this will significantly quieter and cooler for the same level of performance
Will: [00:28:04] It also opens the door I think for water cooling and over clocking all sorts of crazy stuff which is interesting although because of the board design is so dramatically different than it has been in the past I my guess is that the hardware for water cooling will lag a little bit necessarily but I don't know for sure Um Okay I feel like we talked about the Archie cores we probably talked about the tensor cores enough to maybe move on Do we feel like
Brad: [00:28:29] Probably let me see I'm looking for them We have our we have our condensed show notes that I'm looking at And then I'm going back and looking at our raw notes that we just wrote in the presentations It's all over the place
Will: [00:28:40] I just just as
Brad: [00:28:41] here's I'm just gonna read from the tensor cores section of the raw notes bullet 1 This part is magic bullet point two sparse deep learning Bullet 3 start with full weights train it to make a dense matrix and then they prune out the least important weights Then retrain it to remove the dependency on those steps The next bullet point still accurate but less intense Next I don't understand this
Will: [00:29:07] Yeah that's about right Let's use That seems exactly right
Brad: [00:29:09] The last the last bullet point in this giant list on the tensor cores is just I think it's a lot it's a lot
Will: [00:29:17] Um so uh shader core is there's a shitload of them Now they can run like I said independently of the of the Ray tracing core is now I think it used to be that they ran some of the math for the for the ray tracing setup on the shader cores And that seems like it's not the case anymore Um I I yeah I think we should just uh Oh And then the other thing that they added Okay So we have options We can this is I'm gonna choose your own adventure this Brad do you wanna talk about cooler and board design Do you want to talk about the IO stuff The direct storage stuff Do we want to talk about reflex the
Brad: [00:29:57] the reflux stuff for a little later Um have we gotten through all those when we talked about the ray tracing and the shader stuff the tensor
Will: [00:30:04] I think so
Brad: [00:30:05] main stuff with the tensor I mean there's more acceleration around like DLSS and stuff like that I'm not sure where that should fit in here
Will: [00:30:13] Uh well there's the DLSS section down below Uh Oh another thing on the ray tracing If you if you have a 2,000 series card and you've played like control and you remember sitting there watching the shadows and the lights be kind of speckledy like there's almost like gosh and noise on them
Brad: [00:30:29] seen videos of breaking down those artifacts
Will: [00:30:32] So now the de noising on that The de noising on the stuff that the the Archie cores produced will be better because the shader because of the memory speed is way faster The catch speed is way faster on this with the GDDR six X in the shaders can run concurrently with the RT cores now which they couldn't do on touring Apparently I didn't know that but
Brad: [00:30:52] He's going to clean up the output of that stuff and make it look more natural
Will: [00:30:55] my guess is it's still going to be a little bit noisy but you'll but it'll be an improvement over the 2000 series
Brad: [00:30:59] Yeah that's the type of thing That's just going to slowly go away over the course of generations Right As they as they are able to sample more, casts more rays and so forth
Will: [00:31:09] yeah that that stuff happens because they're casting a relatively like if you're making an animated movie like a Pixar movie using the same kind of Ray tracing they would cast millions of rays And on this they're casting thousands of raise probably
Brad: [00:31:21] and just kind of interrelating the results in between
Will: [00:31:24] well so they yeah that's exactly it They're usually they're catching a whole boatload of rays and then they're using like a upscaling to fill in the holes between the rays that landed And that produces a lot of noise right now
Brad: [00:31:36] to be clear because because casting enough res to get an accurate result would be prohibitive Computationally is what you're saying That's just that's too much for the hardware to handle or they're kind of fudging it They're there they're doing a low precision operation and then sort of fudging
Will: [00:31:49] It's Yeah it's funny because one of the things like in the old days when you pre-baked lighting like in Quake or I guess unreal pre-race old unreal doesn't but new Anyway so we fudge the lighting because it was too computationally expensive to do dynamic lights So then we started out as adding some dynamic lights and we started doing light probes and stuff like that But um Now we can do the Ray trace lighting the real the real lighting the true lighting But we have to fudge that because we can't do it fast We can't do enough of it So we're fudging the results of that rather than fudging the the results of the it's like one layer less abstract Now I guess
Brad: Just a little less fudge than before
Will: So one of the other interesting things that came out as a result of all this is that the workflow and the performance like the performance limits are dramatically different Like on a on a pre-baked on a dynamic lighting scene using more traditional rendering methods Like the number of light sources is a real game and it is still here Like the complexity of the math goes up exponentially Every time you add another light Um but The interesting thing is that like the poly density of the scene when it's Ray traced doesn't impact performance as much And that I didn't really understand because it's all trying like you're you're just looking at surfaces So um whether there's infinite surfaces or relatively small number of surfaces it doesn't change that the fact that the math is the performance is gated by the number of lights in the same with the rate ray-traced graphics
Brad: [00:33:27] the upshot like I don't know how much we're getting lost in the weeds here and how much people care about all this stuff that like again I just want to reiterate like you know when they come out and say like Oh this one is two times faster than the last one Like that sounds a bit bold you know like that's that sounds like kind of a grandiose marketing driven sort of claim But see but seeing all this stuff on the scenes really did give context of that stuff in a way that made it it made it easy to believe that they are seeing numbers You know obviously those are in ideal cases Like it's not going to be like that across the board but it at least they at least backed up their claims in the initial stream in a lot of ways
Will: [00:33:59] We’ll when the benchmarks come out like when they're there will be independent people running benchmarks in in uh in days or weeks before the cards are available.
Brad: [00:34:05] Yeah but at least for me coming from you know coming from the games press where they like companies tend to shy away from technical stuff as much as possible and make everything digestible
Will: [00:34:14] Yeah How how how was this day for you compared to like Hey we want you to come and play the new gears game
Brad: [00:34:20] well like I said this was this was this was a very eyeopening look into how intensely technical this stuff can get Or I was just I was shocked to see this level of detail exposed to the press but obviously like you know there there are a lot of Tech sites out there for whom this is their bread and butter right
Will: [00:34:35] Yeah Well I mean I guess yeah like when you go to a games thing but I've been to games things before it's not like you get 12 hours on how borderlands three works and then they send you home with a copy of the game to try it out I guess Um
Brad: [00:34:49] Well whatever it's it's media criticism versus sort of you know technical writing
Will: [00:34:53] like technical writing Yeah exactly Um so okay Reflex is this is this new thing that they're rolling out They they said that basically over the last couple of years they've been looking at the impact of latency on on performance specifically for they started with eSports people but they've been looking at more normal human beings now Um and and latency Uh in this case means the time between when you click the mouse and when the thing that you wanted to happen by clicking the mouse happens on screen Um typically it's what a lot of people call input lag Uh they I feel like they kind of didn't like that idea
Brad: [00:35:31] are They're actively campaigning against the use of the phrase input lag They seem to want to move to something that they refer to as system latency which
Will: [00:35:38] is
Brad: [00:35:39] kind of the whole the whole stack of latency from from physically making you know clicking a button creating input to like through the graphics pipeline to X Y and Z until it is reflect on There's a lot more steps in there is what they’re saying.
Will: [00:35:52] It's like it's there there's a there's there's like the mouse there's the PC you know USB bus CPU memory all of that stuff the stuff that's happening and like what the game is doing And then there's the render queue out through the graphics pipeline and onto the monitor So each of those things can add latency
Brad: [00:36:09] And that that might that might come off as overly pedantic But I think the point they're making is like Hey there are a lot of different phases in this process and we can optimize different phases in different ways to try to try to minimize the whole thing
Will: [00:36:21] And also this is the other thing that I was surprised by is that uh so they said they said typical latency The system latency is around 38 milliseconds uh before they started working on this stuff Uh they're they're finding that uh 20 to 25 milliseconds is like what a pro gamer somebody who's relatively young can handle in terms of physiological reaction time Um and that I'm sorry Average fit system latency was 55 milliseconds um
Brad: [00:36:52]So higher than even the average person's reaction time.
Will: [00:36:54] higher than the average person's reaction time Uh the thing the thing that uh So this helps with what's called peaker's advantage And this is something I was aware of but hadn't really thought about and hadn't thought about a good way to metaphorized before But um I think it's easier The idea is peaker's advantage means that if you are if somebody is watching a corner and you peek the corner you have The time you'll know what's on the other side of the corner before the person who's the peaker will see the person who's camping in the corner before the person camping in the corner and get the signal that the person has peaked And it's a necessary function of the physics of the time the tie the 38 milliseconds or 55 milliseconds It takes the computer to render all that stuff after it gets the signal Is is that thing And and the important thing to note about latencies that increasing frame rate does not necessarily decrease latency depending on how you increase your frame rate So like you see CS go and Overwatch and Valorant and people who are playing those really low system impact really high frame rate games at 500 frames a second a thousand frames a second and it may not necessarily be helping them reduce the latency which is the time between the mouse click and the and the action taking place on screen
Brad: [00:38:11] was this a was this phenomenon being reflected When I played trials of the nine with a hardcore destiny crew And every time we would go or every time they would go around corners they would fire around the corner as they were looking
Will: [00:38:24] yeah The pre fire
Brad: [00:38:25] yes they were pre-firing they were
Will: [00:38:26] Yes pre-firing
Brad: [00:38:28] firing before they even knew if somebody was there or not
Will: [00:38:30] Pre firing is how you counter a peaker's advantage Now this is Peter's advantage is independent of lag too of like of like network lag So those are two different things And and like I feel like peaker's advantage probably is usually out superseded by network lag on games that you're playing on the internet but probably not If you're at like a competitive tournament or something like that on the land where there's lower latency or if you're playing like you have to have an incredibly low ping Um so the upshot of the reflux stuff is that they're able to get the latency down a significant amount like to 12 milliseconds in some cases um previously the way you measured this was by setting up a high speed camera in front of a monitor and wiring a mouse It the wiring the mouse switch into a trigger for the high speed cameras so that you can see lot when all this stuff happens You literally just count frames Uh and uh it's a setup that nobody can replicate at home The camera those cameras start at like 10 or 10 10 or 15 grand Um and it it was it was time consuming and expensive Uh they've built hardware into the high end But it seems like the new high end G sync monitors If they say they have reflux support they have a piece of hardware in there that you plug in the mouse into a special USB port
Brad: [00:39:44] it has to be it has to be the specific marked USB port I thought that was amusing
Will: [00:39:49] Well and then you use the onscreen display on the monitor to show where the crosshair like where the shooti spot is on your monitor Um and you hit the reflux button and you fire the mouse the monitor detects when the pixels when the pixels change on the screen and the delicate in the designated area And then You get a thing that pops up in Geforce experience It says Hey here's what if you have a supported mouse
Brad: [00:40:15] That was, I was going to say like let's do it We have to be clear It's not not only do you need a supported monitor but also a supported mouse.
Will: [00:40:21] Well so you can get the total soup to nuts latency without the supported mouse But if you have a supported mouse it'll break out your mouse latency too
Brad: [00:40:28] Okay That's just for extra data
Will: [00:40:29] Yeah That's
Brad: [00:40:29] extra extra profiling
Will: [00:40:31] Um so so the upshot is you can find out that your mouse takes four milliseconds and you know there's five five milliseconds of CPU light latency and then the video card and the monitor handle nine milliseconds and you get a 17 millisecond total or whatever those numbers just picking arbitrary numbers there I apologize Um I think this is really neat I'm really really curious to see it Like they said I specifically asked if this impacts normal human beings as well as like shrouds and liquids and all those like aim monsters, And they said that they've they've seen more impact on not on casual gamers because a lot of what the pros do is train themselves to buy to like ameliorate Things like peakers late peakers advantage by pre firing and latency uh compensation by like just knowing how to aim better Um so they're they're they're developing muscle memory to counteract these things whereas you or I would probably be reacting to what we see on the screen
Brad: [00:41:36] Right This seems a reduced version of this seems good for the average consumer in terms of the hardware demands like like how many people you know obviously there are going to be people in the consumer space that always want the best of the best and we'll go out and get these supported mouse in the supported monitor and et cetera et cetera But like it's a tall ask for the average person to buy all these peripherals that specifically support this feature to get all this stuff
Will: [00:41:58] Of course but I mean if it comes in just your G sync monitor then you're you're good to go
Brad: [00:42:01] Yes Yeah So if they start building it into the your average monitor and you're not having to buy an $800 monitor just to get it
Will: [00:42:07] So um the other thing that's nice is that all of this stuff is going to like this applies not just to amp your cards but to pretty much everything that they said was currently in the market Like the people are using right now.
Brad: [00:42:18] So you're you're saying like if you get the monitor in the mouse it'll work with your 2080
Will: [00:42:22] Well not only that but like the reflex stuff that reduces the latency without having them like you can get the benefit of the latency reduction on any card on whatever monitor you're using They may have to be g-sync monitors I didn't ask I didn't ask the follow up there but um
Brad: [00:42:37] I think you and I both had the same thought during this part of the presentation which was that it almost seems surprising that there isn't a USB port on the graphics card at this point So
Will: [00:42:46] there was on the ampere cards and on the on the turing cards like my 2080 has a USB-C port
Brad: [00:42:52] huh Well but it is that not for less, That's for display out Isn't it?
Will: [00:42:56] well it's for display out and for a VR headsets but it's basically just a normal ass Like is it normal You can plug your mouse into it if you want
Brad: [00:43:03] I that's that's one of those things where USB USB being a physical connector is just a black box for all the things that might be able to do like I always think of it as like Oh well if it's on a video card it's probably just a video output but in fact it could probably do everything a USBC port can do which is functionally infinite That's confused That's confusing But but yeah like I it does make me and we'll get to this more with the IO the storage stuff that they're doing which I probably gonna get to you next But it really starts making me feel like the graphics card is more and more becoming the heart of the PC
Will: [00:43:35] Well if if you think if you're if you play games Think about how often you upgrade your CPU versus how often you upgraded your video card Right And like in that if you're using if you're thinking about which is more important I spend more I've I've upgraded my graphics card probably two or three times for every CPU upgrade that I’ve done in the last 20 years.
Brad: [00:43:52] Yeah That's that's that's not yeah that's been true for quite some time but I'm talking about more of the the number of functions that are being moved on to the graphics card
Will: [00:44:01] yeah the thing that the thing that's happened since direct X three is that anything a massively parallel computing device can do better than a serial computing device would even like an eight core You know if you have an eight core CPU with 16 threads you can do 16 things at once Theoretically right The GPU has 10,000 shader cores.
Brad: [00:44:21] Sure
Will: [00:44:21] So it can do like it can't do 10,000 super complicated things at once But if you can break down 10,000 complicated problems into the a hundred thousand really small steps it can do those really really fast like insanely fast.
Brad: [00:44:35] Yeah I mean you're you're right It is a massive amount of compute power Just sitting there waiting to be utilized for things other than other than graphics Right Uh but still it's just it's just a weird evolution of the the form or the the anatomy of a PC right That we're sitting here talking about potentially plugging your mouse straight in your graphics card at some point in the future
Will: [00:44:53] I yeah I I so I was when they said you had to play the mouse into the monitor I was like Oh okay That makes more sense But like I wouldn't have been surprised if uh if I like well it won't surprise me too much If at some point in the future we're plugging like SSDs into the your your if the video card has an NVME slot on it for reasons that we'll get to in just a minute um Yeah it's a weird this is a weird like so the next thing is the IO improvements We've talked about this a little bit in the big brief but they really got into it on on Thursday
Brad: [00:45:26] And this the demonstrations were the breakdowns they gave you were super helpful for visualizing why this matters what it does and why it matters
Will: [00:45:33] Yeah So so the upshot is right now when you install a game and you're you run it and you're copying stuff over you're copying all the textures and the assets that make up the game from your SSD or God forbid your hard drive
Brad: [00:45:47]at runtime to be clear you're not you're not talking about and you're not talking about and when you install the game
Will: [00:45:50] I'm not talking about downloading from steam
Brad: [00:45:52] this is loading the game when you're playing it
Will: [00:45:55] Yeah So when you do that uh it comes in and it goes from the SSD across the PCIE bus to the CPU into system memory where some math happens and it gets decompressed And then these huge decompressed assets go back through the CPU onto the PCI express bus into the GPU and into GPU memory
Brad: [00:46:13] it's a it's a very winding process for that data to go from the compressed installed to to the graphics card where it needs to be
Will: [00:46:21] Yeah the uh and also there's a lot of bottlenecks And maybe if you're doing say hundreds of thousands of really small IO operations the CPU because it can only do eight or 16 at a time is slow for that Um What this what this is there's an API called direct storage That's a part of DX 12 is coming out next year Uh lets you bypass it It's uh I think I feel like this was probably built for the Xbox series X
Brad: [00:46:48] yes I don't I think it might be the exact same middleware I think it's the same
Will: [00:46:52] It is the same API Yeah They said it's the same API but I I yeah I feel like this was built to solve a problem with I mean so the the consults are a little bit different since they use a unified memory bus Um so you know there's there's system memory and GPU memory are the same memory on the Xbox and the PS5.
Brad: [00:47:10] yeah that's been true since I think the 360
Will: [00:47:13] The definitely the 360 I don't know I don't know how it worked on the PS3 I can't remember But but yeah the AP all of the APU based consoles have had shared memory
Brad: [00:47:21] The PS3 was weird I know you probably don't care but at the Guthrie was split down the middle I think it was
Will: [00:47:25] Oh that's right
Brad: [00:47:26] 512 Meg half for the GPU which I think was added to that system afterthought And they either have for system memory
Will: [00:47:34] I believe that was added at the request of Tim Sweeney who is the only person in the world who has the hoist to get new video or Sony or Microsoft to do add hardware
Brad: [00:47:43] Where whereas I think it was flexible on the Xbox So you could kind of define how much was what the not that I mean we're kind of going to get off On a tangent here but uh I was always in under the impression that that was a unilateral advantage for consoles to just have one giant pool of memory that they could kind of use flexibly however they want But there was a I forget it was a Tommasi or I
Will: [00:48:04] Tony Tommasi talked about this.
Brad: [00:48:05] Who gave who you gave the breakdown the system memory and graphics memory have different priorities That's that Um what is it It's latency system memory latency is the thing you want to optimize for and graphics memory it's throughput Is that right I don't know if we I don't know if we wrote that stuff down or not
Will: [00:48:21] I think that's right I think yeah So they're talking about raw um cause in graphics traditionally the speed of system memory dependent on how fast GPU memory dependent on was directly tied to how fast you could render if you're doing Rastro based rendering um with with um So non Ray traced old school you know shaded textures being lit by texture maps and stuff like that Um
Brad: [00:48:47] But this this was unresponsive Somebody somebody in the audience asked a question cause they were talking about the new memory on these cards So they were like if this memory is so great why don't PCs Just use it for everything Why aren't you also putting you know dims of this stuff into your motherboard and and uh I don't know if that ties into the the direct storage stuff so much where you know assets are going through the system memory before they go back to the graphics card I guess it would right I'm sorry I don't know We're getting way
Will: [00:49:14] No no no Say that again
Brad: [00:49:15] pretty far a field of the the direct storage stuff here Um I don't know I said what I was saying is I don't know if that dichotomy is relevant here where you know like the problem is that all these assets are being shunted through system memory before they get to the graphics memory where they actually belong And if that if that memory is not optimized for throughput like he was saying
Will: [00:49:35] Well it is optimized for throughput but in a in a much much less it's much more they care much more about latency since
Brad: [00:49:41] what I
Will: [00:49:41] like the goal of system memory is to keep the CPU fed at any given moment And the goal of graphics memory is to get shit in and out as fast as possible Cause they're doing whatever they're doing 150 times a second
Brad: [00:49:50] Sorry I know I'm I'm just gonna ramble here but
Will: [00:49:52] no no You're good
Brad: [00:49:53] I find this stuff I find this stuff fascinating Even if I'm playing it out in the clearest fashion
Will: [00:49:57] I was trying to find the Tony's quote I thought I'd saved it but I didn't about the GPU and the met shared memory pool on the consoles And I don't think I saved it I did not Um anyway yeah so so this is the neat thing about this is you know where we went from SSD to PCI express to CPU to system memory to decompressing to CPU to PCI express to GPU to GPU memory
Brad: [00:50:23] so it's a long process
Will: [00:50:25] I mean look it's still happens millions of times a second but yeah now they basically the games if they support the direct story API can load assets straight from the SSD to PCI express across PCI express to the GPU bypassing the CPU entirely And this is important because NVME SSDs are optimized for tons and tons and tons tens of thousands of sequential reads and writes
Brad: [00:50:50] they are massively parallel
Will: [00:50:51] Yeah And in a way that sata SSDs and um and especially hard drives were absolutely not So so theoretically if you have a game that uses direct storage and can write directly to the GPU , GPU memory or read directly to the GPU and GPU memory you should be able to saturate the this is going to move the bottleneck from the CPU
Brad: [00:51:14] to the storage device
Will: [00:51:16] to the storage bus to the to the to the actual capacity of your storage
Brad: [00:51:19] you're basically able to process and load the data as fast as you can read it off that fast ass drive as opposed to it hitting this super small bottleneck on the CPU
Will: [00:51:27] Yeah Uh the the the downside of this is the games do have to support the new PA API to see the benefit And since the API is not out yet it's probably going to be awhile before we see it the games
Brad: [00:51:37] Yeah Uh another another good question I thought that came out of the audience during this section was um somebody asked since most games are GPU bound Why would you want to move more of this work onto the GPU instead of just letting the CPU continue to handle it Uh because theoretically the CPU is sitting there with some amount of idle time Um they have their their answer was basically like you know we think it's good to give developers a choice to use this They obviously don't have to And for some games it may not make sense to actually use this to you And they might want to stick with the old way of doing things but but it is an option for games where it does make sense
Will: [00:52:11] The thing the thing that I took away from this is well there's two things One is that games that previously would have had discreet level loads If you can rely on this technology being prevalent which in the PC is going to be a long time before you can um Then you may be able to do things like streaming level loads and have games be more open rather than having to do the the you know the run through the S curve and and and have a load in between there half-life style Like we've been doing for the last 15 20 years
Brad: [00:52:39] Yeah this this was a did you watch the Mark Sarney presentation on the PS five months ago Like this this is this is the exact same thing he was talking about in the PS five Like this was another nice illustration of why this stuff matters and how they're accomplishing it You know same situation They're have they have they have dedicated hardware to handle this IO and the decompression of assets And this is kind of the same thing It's just it's on the GPU in this case
Will: [00:53:03] Well and the interesting thing I wonder if on the PS five they're putting this hardware there they're implementing this API on the storage side rather than on the GPU side So so what are the other things that they talked about in this part that I thought was fascinating is they were loading this up on thread rippers which had 48 threads available And like even then the GPU was just romping the 48 core thread ripper
Brad: [00:53:27] so I've got.
Will: [00:53:27] in terms of in terms of level loads
Brad: [00:53:29] So I've got um I've got I saved that slide here So I've got those numbers
Will: [00:53:33] Oh yeah let's have ‘em.
Brad: [00:53:34] me let me try to summarize here because they didn't they didn't expose some of the numbers you actually want and I need to dig for them but basically um uncompressed assets loading off This was for their a what do they call this tech demo The
Will: [00:53:47] The marbles thing
Brad: [00:53:48] one that they showed that I got I I'd a I want to get my hands on one of these cards and be I really hope they put out that tech demo because I really want to play with that thing That's like
Will: [00:53:56] They usually they usually do
Brad: [00:53:58] the nicest looking um tech demo I've ever seen But
Will: [00:54:01] It was pretty good
Brad: [00:54:02] um so loading demo in or loading assets from that marbles tech demo uh uncompressed off of a hard drive over a minute 62 seconds off of off of an NVME a PCIE SSD But with CPU decompression which is the old way of doing things right 5.2 five seconds
Will: That’s a pretty good improvement.
Brad: uh off of that same NVME drive But with using the new GPU decompression for the assets it was 1.6 two seconds
Will: [00:54:30] And that was on the 32 That was on the 24 core thread ripper so 48 Threads
Brad: [00:54:33] yes So yes, Did you say 24 cores
Will: [00:54:38] 24 cores 48 threads
Brad: [00:54:39] yeah he's 24 core thread ripper which is a CPU that not most people have in a gaming machine still could only do that in the over five seconds where it only took a second and a half with this GPU
Will: [00:54:49] I wrote a note at this point cause this is the moment at which they would usually if if they weren't really confident in this technology they would do the demo on like a 6,600 K or something like an eight year old quad core And uh the fact that the that the new 24 core got swamped is a very telling
Brad: [00:55:07] Yes this was not this was not here's how it stacks up against the CPU Most people have from five years ago It's it's here's how this stacks up against the top of the line CPU on the market
Will: [00:55:16] Yeah so
Brad: [00:55:16] I I'm sorry No I was well you go ahead I'll I just was going to summarize something
Will: [00:55:22] I just wrote in here what the last note in this section is shit is about to get really bad for PC gamers who don't have SSDs yet Um
Brad: [00:55:30] Yeah.. I don't That is maybe the biggest open question to me of how games are going to go moving on the end of this new console generation where the games on those consoles are being designed around having this as a as a baseline
Will: [00:55:41] like can you port ratchet and clank warping from world to world through those portals to PC I don't I I maybe but it's the market can be big enough to support that game Probably not
Brad: [00:55:51] I mean that's a you know that's a hypothetical example cause there's no guarantee that game will come to PC Even even with Sony now porting some of their first party games that that one may well never come over but there are going to be a ton of third party games that absolutely will get ported and and how those get ported in and who gets who gets shut out based on what kind of storage device they have is going to be a strange situation
Will: [00:56:12] one of the questions I have for people who are making games is how I haven't looked at the developer docs for this API specifically but there's a real like how if you choose to build a direct a direct storage game how How do you is there a graceful fall over too for people who don't have the hardware that supports direct storage and also you know something nobody asked and I didn't think to ask on the day is is is this an ampere feature or is this going to come to the whole Nvidia line Like is this something that they can back back port over to the older cards
Brad: [00:56:50] th they in every other example where that was the case they came out and said it explicitly So I would assume they would have done the same here if that was true And this also to me and I'm just guessing but this sounds like something that relies on specific hardware in this card
Will: [00:57:02] I this seems like general math stuff You could run on tensor on on a on tensor shader corse
Brad: [00:57:07] you might be right.
Will: [00:57:08] anyway um
Brad: [00:57:10] like the game design stuff I mean I don't know how much you want to dwell on this but like it it really goes to the fundamental question of game design and how much you want to rely on specific hardware features Because like again in that um in that Cerny talk he was he was talking about their IO stack being fast enough to load assets faster than you can spin the camera around in a scene and a game so that you literally can be loading more assets as you look around in a room And and and but like that is that is a feature that you absolutely cannot have in a game that has to run off of hard drive or even a sata3 SSD. So like the the most inclusive strategy but also like in a way the kind of I don't know if sad is the right word I want to use but you'd be leaving a lot on the table if a lot of third parties ops to just not design games around having that kind of feature because they need to come to slower devices
Will: [00:57:56] I mean the one thing that will happen is that these consoles will last somewhere between five and 10 years Right Like the last couple of generations have both gone at last 7
Brad: [00:58:04] seven About seven
Will: [00:58:05] Yeah
Brad: [00:58:06] Seven eight
Will: [00:58:07] so If you assume that the clock on that is starting now by the time we get to the horizon zero Dawn is of this generation
Brad: [00:58:14] there'll be pretty widespread
Will: [00:58:16] be enough of a PC market to support this
Brad: [00:58:18] you're right You're right It's the prices on that stuff are falling so fast Like I saw that uh like a 512 gig uh NVME PCIE three drive now Like I saw one on sale at Western digital for like $65 the other day And I pay I paid over $200 for one of those two years ago So
Will: [00:58:36] two terabyte Decent NVME SSDs are now priced about what the one terabyte drives were this time last
Brad: [00:58:43] Yeah And to be fair I think those the the low end ones are not quite in the same ballpark of throughput as the console ones but they're getting into that territory
Will: [00:58:50] I mean there's time
Brad: [00:58:51] Yeah you're right So like by the time these games are prolific maybe this will be kind of a solved problem in the PC space because that will be just commodity hardware at that point
Will: [00:58:59] I mean I do think in the short term we're going to see especially people who are focused on third party unless there's an easy fallback that doesn't require redesigning the actual like levels and assets in the game We'll see Probably people are opting for the for the least common denominator which will be not not using the stuff on third party titles
Brad: [00:59:20] Yeah Uh on the uh
Will: [00:59:21] or forking the I mean it's entirely I mean maybe I'm nuts Maybe people will fork the assets and just build to like two versions of transition areas for xbox, ps5 and PC.
Brad: [00:59:35] I was absolutely not unheard of back in like the you know the PS2 original X-Box GameCube days of games games basically being the same across generations or across platforms except for
Will: [00:59:43] Like the level loads and
Brad: [00:59:44] designed around specific limitations here and there Um
Will: [00:59:48] adds a lot of time to the development and testing and stuff which is a bummer
Brad: [00:59:49] yes Yes That is
Brad: [00:59:50] uh that is a ton more work for the developer obviously Um on the flip side the other thing I wanted to say about this is I have seen I don't know about you I don't how much you read like gaming message boards and stuff but there has been so much skepticism around the SSDs and the new consoles and and how much of an impact they're actually gonna make And I think that is very much coming from you know that transition where We were all on SATA3 hard drives and you know the first Sata3 where we on SATA2 when the first SSDs were coming out I can't remember
Will: We were on Sata 2 yeah
Brad: Right So it's been it's been a lot of years
Will: It’s been 10 years
Brad: right But it's been a lot of years of SSDs being shoved onto a bus that was not really appropriate for them for the medium Right
Will: [01:00:31] I mean the thing that happened is that the first SSDs were so regardless of bus speed were so much faster than than this is like the latency versus frame rate thing Again the first SSD is seek times were so fast compared to traditional hard drives that they felt they felt infinitely faster So there was a huge performance bump when you got your first SSD in a computer and then the performance bump from SSD one SSD two probably didn't feel like anything you didn't notice
Brad: [01:01:00] That’s kind of what I’m getting at the floor the floor for how much uh loading times end games could fall on SSDs has been kind of artificially high for a while You know and I've done some I've done a little bit of testing on SSD load times with games in the last couple of years And it's like you know you see situations where it's like okay it's like seven or eight seconds to load this instead of 15 from one generation of SSD to the next something like that You know And I think a lot of people look at that and think like yes as these are faster but they're not like instant loading fast you know because they're they're looking at this kind of old model Where the devices were constrained The games also were constrained right Like things have not been built around this like radically different way of of living assets
Will: [01:01:38] Well and and on the PC I think I want to say red dead was the first thing I saw that had like you should have an SSD to play this game Um
Brad: [01:01:46] Even stuff like WOW Shadowlands the new expansion uh they actually they list SSD under their minimum requirements for that expansion I don't know if that literally means you have to run off an SSD or not but
Will: [01:01:57] But but I mean you have to remember all of the current gen consoles have hard drives right So like anything that was designed to ship on PS four and Xbox one is is designed to run off of a spinning platter with sequential with a relatively small number of sequential reads
Brad: [01:02:14] Right Yeah All I'm getting at is that there you know there is there is a lot of this existing sales skepticism that these new SSDs are going to actually matter more than the oldest SSDs have for the last five to seven years And seeing seeing stuff and seeing this broken down in a technical way and also seeing that marbles demo and the times there you know Like you know a second a second and a half for a load that takes even compressed Was it like an uncompressed over a minute on a hard drive with compressed assets still like 35 seconds or something
Will: [01:02:42] Well and the the fun thing about using compressed assets is it increases your actual throughput because the video card sees the compressed assets is as full size So you're transferring if you're transferring an asset that's compressed two times or four times then you're literally seeing four X the actual throughput of the PCI express press if you’re saturating the PCI express like it's Anyway, this is cool.
Brad: [01:03:05] Yeah Yeah I know I know we're getting way into the weeds here and not not to not to belabor the point but there are there are made there are major optimizations here that are gonna make a huge difference that people just have not had time to see in real in real world situations yet
Will: [01:03:17] So if you're skeptical about the value of SSDs
Brad: [01:03:21] or specifically specifically the next gen SSDs that run on the PCIE bus
Will: [01:03:25] yeah They're like this is this is something I believe let's put it that way Um yeah this is not this is not bullshit marketing stuff This
Brad: [01:03:36] Yeah It's all there It's all there on paper And the technical way is what I'm saying Like the the the the mechanics of this are well understood
Will: [01:03:44] Uh next up let's do DLSS cause this is cool And I didn't really understand this stuff particularly well before Um but this is a DLSS is short for deep learning super scaling which is a marketing term Uh but basically what it what it is set up to do is to take games that can't run at 4k at a playable frame rate rendered them internally at a lower resolution and then applies machine learning to fill in the data and let it run it Say 4k 60 when it would normally run it 1440 60 or something like that.
Brad: [01:04:14] So I mean is it is it accurate to say that this is just a very fancy AI driven form of anti-aliasing or upscaling
Will: [01:04:21] Kind of yeah So like uh if you think about how AA works like super scaling super sampling works by rendering the whole frame at like some resolution that's higher than what you're going to show on the screen And then it averages a bunch of pixels It does a rolling average of the pixels in like whatever size grid To get to turn like four pixels or eight pixels or 16 pixels into one pixel smooth out all of those all of those jagged edges Basically what this does is just does that but without doing the scaling back down so that it stays at the high resolution and you end up with Um they uh they they train neural nets on actual frames from games that are matched with perfect frames rendered either with super sampling or with accumulation rendering accumulation renderings where they render each frame like five times and blend that that uh blend the results to get like one high quality result out of multiple passes
Brad: [01:05:14] So this is the part where I'm eyes start crossing trying to Parse all this where are where all the where where's the dataset coming from that they were using to train this model Like where are these frames coming from Like is this this is not something you're not talking about The frames that are being rendered in real time in the game Right And like is this something that has happened
Will: I am
Brad: Wait Okay
Will: [01:05:31] So so when they're training they render the games in a controlled situation so that they can get exact duplicates of the same frame Right So like they'll run they'll have a loop that they run the game through That's like running say a death Stranding is a is a good example Cause they'll have like They'll run the characters through every aspect of the game so they can see all parts of the game Right And they'll record those play throughs And then they'll render them back both on normal desktop hardware with the actual frames and then also on the supercomputer at really really really really really high resolutions up to 8k now.
Brad: [01:06:05] so it's okay Let me distinguish here So what I meant was the frames that are being generated by your game on your machine as you were playing that those that is not it's not training itself in real time as it's also out That's
Will: [01:06:16] no no no no no
Brad: [01:06:17] what so the training the training is taking place like at Nvidia prior to this this being put into practice
Will: [01:06:22] this is this is why games have to actively support DLSS
Brad: [01:06:27] Right So that that that was that was the natural conclusion is that every game needs specific support for this Cause they have to do the work per game
Will: [01:06:33] Yes It's Yes So um so your training it against the perfect frames And then on the card the tensor cores the things that do all the machine learning acceleration which normally aren't being used while you're playing most games can can run on the frame buffer the output of the frame buffer and scale it up using the data from this neural net that they've that they've trained
Brad: [01:06:53] That's insane
Will: [01:06:54] Um and the examples that they show always look like magic I have seen I have seen a wide range of wow this is really amazing too Huh That's a weird artifact where like some stuff that pops up this shouldn't be there as popping up in the in the frame because you know machine learning is inherently goofy
Brad: [01:07:11] I that’s what I was going to say that is the story of machine learning writ large right Like it's like like incredibly like amazing almost magical effects with just the occasional like like mind boggling aberration right Like something
Will: [01:07:22] why is there a dog face in that stop sign That's really weird.
Brad: [01:07:24] it's like it's 95% and magically in insane and amazing and impressive and like 5% like awful and completely wrong
Will: [01:07:32] We have a channel in our in my work Slack called nightmare factory that's for places where the machine learning does bad things Um so so yeah you basically I've used that too Play death stranding at 4k 60 on a 2080TI It works shockingly Well it requires a lot of work on the part of the developer They've now added support for greater than 4k resolutions So previously was limited to 4k at the at the peak Uh now you can go up to eight K which is how they had um they had some people come in and look at uh eight K They had somebody will come and look at games on 8K screens I believe they were giant like 85 inch OLEDs Um but uh but yeah it's like basically the benefit of this is that you can get Really good looking high image quality at resolutions that would be impossible to hit on existing hardware maybe for the foreseeable future um by by using AI machine learning to scale up the resolution of the rendered images which is just sounds goofy
Brad: [01:08:42] I didn't know Those screens were available in a form that could be moved around and hooked up like normal I assume that that stuff was all like still pretty well I guess they've been showing 8K screens at CES for a year or two now
Will: [01:08:52] you can go If you want to spend $50,000 you can buy an 8K TV right now Brad
Brad: [01:08:56] Oh I didn't know they were actually on the market
Will: [01:08:58] Yeah Oh yeah So um while we're on that they did add support infrastructure support for 8K 60 Hertz gaming Uh the these card support HDMI 2.1 which has bandwidth to do everything on one cable Previously 8K screens were either two display ports or four HDMI 2.0s I think um Uh they also added support for AV one The new video codex is one of the first eight K codex There's no support for H.266 yet in these cards but that presumably will be in next generation or maybe in a feature bump
Brad: [01:09:32] Yeah That just got that just got ratified a month or two ago
Will: [01:09:35] I don't think any of this stuff matters for normal people
Brad: [01:09:38] Probably not
Will: [01:09:39] Yeah I think this is like for for tech demos and you know weird stuff
Brad: [01:09:43] I had not even heard of Av one I'm not familiar with that
Will: [01:09:46] So uh it's there there've been two codecs for um blu Ray that's there's there's an AV predecessor And then there's
Brad: [01:09:55] I see Yes This says
Will: [01:09:56] H.264 as well.
Brad: [01:09:57] As the successor to VP nine which I know I know that's what YouTube uses for their four four K videos
Will: [01:10:03] I don't know if they do anymore I think they're using um
Brad: [01:10:06] or they did for a while At least
Will: [01:10:08] his his VP nine the ogg vorbis follow
Brad: [01:10:11] I believe so
Will: [01:10:13] Anyway
Brad: [01:10:13] I think that's right
Will: [01:10:14] people are screaming at their radios right now Um yeah I I like realistically are you going to play 8K games on are you going to have an 8K monitor in the time that you're going to have this video card Probably not
Brad: [01:10:26] I don’t know for a fact that I'll have a 4K monitor in the time that I'm using this video card
Will: [01:10:31] Well I mean I think the big takeaway is that with the 3090 4k is probably possible
Brad: [01:10:37] totally possible It's it's more it's more a matter of waiting for the monitors to catch up Honestly like there just aren't a lot of like high refresh 4k has just not been a thing really And a lot of that has been limited by the connection types the cable types
Will: [01:10:49] Well and and by the panel speeds
Brad: [01:10:51] Yes So um
Will: [01:10:54] Um let's talk
Brad: [01:10:54] I might be good on 1440 P for a bit
Will: [01:10:57] do you wanna talk about the coolers real quick
Brad: [01:10:59] Yeah That stuff seemed pretty impressive I mean I don't know a lot about cohort design I mean obviously I know what a well put it this way I know what a heat pipe is I'd still don't know internally what a vapor chamber does
Will: [01:11:08] Okay So cooler design is that they the big thing you noticed when they showed the board cutout is that there's a whole like the end of the board is a V-shaped to accommodate the second fan
Brad: [01:11:17] the the PCB of this video card looks metal as fuck It's like a it's like a spike V it's almost like a Klingon on weapon or something
Will: [01:11:24] Yeah
Brad: [01:11:25] I've never seen a PCB that looks like that before It's also incredibly dense
Will: [01:11:28] it is it is So
Brad: [01:11:30] PCB is like half the footprint of the one on the 2080 from the looks of the comparison
Will: [01:11:35] Yeah And it has more memory and like it's it is it is incredible The the density they have memory on both sides of the board
Brad: [01:11:41] like no bare board space
Will: [01:11:43] well yeah There's no yeah
Brad: [01:11:45] with capacitors and memory modules and just yeah
Will: [01:11:48] I have to imagine it was a nightmare to design and get everything crammed in here but they did the big V cutout so that they can have basically half of the board is fins and fan So the there's a cooler that fits into the into that spot That's connected to the heat generating parts of the board the GPU and the memory with a vapor chamber and heat pipes a vapor chamber heat pipes basically same idea They they use uh the magic of Boyle's law which means that Converting things from liquid to gas takes a fuck load of energy compared to raising the temperature of a liquid or a gas one degree Celsius So like going from if we're using water as the example going from 100 degrees Celsius liquid state to 100 degrees Celsius sorry let me do this another way Going from 90 degrees celsius 99 degrees Celsius to a hundred degrees Celsius Liquid state takes like 80 calories per gram of water Right When you want to go from 100 degrees Celsius liquid state to 100 degrees Celsius vapor state It takes 540 calories of energy per gram of water
Brad: [01:12:54] so that that phase or that state change
Will: [01:12:57] Exactly
Brad: [01:12:58] is a huge energy sink
Will: [01:12:59] So all of the cooling apparatus for modern graphics cards has relied on the ability to remove heat by converting things by by basically setting up the state at say state inside the heat pipes and the vapor chamber in such a way that something is the sweet spot for the temperature on the card And the cooler is going to be at the vapor change temperature So so that they can juice that by by changing Anyway it doesn't matter
Brad: [01:13:26] does does that mean that moisture is just constantly being cycled in a way that it converts back and then can just continually be changed from liquid to vapor
Will: [01:13:33] Yeah so the vapor we'll go move to the cool end of the pipe and it'll condense and it'll roll back down into the pipe and it'll get back into the hot part And then it'll just it's just constant convection There's no mechanical process They're usually um the benefit of this is you can move a boatload of heat The blow through fan is interesting because currently in most systems where the video card is plugged into the top PCI express slot like you have air coming in the bottom in front of a case Blowing up theoretically over an air cooler on the CPU or maybe through a radiator out Um what this is doing is is removing like that That video card in the middle of the system created a giant Edie for air flow where it basically stalled or maybe got sucked into the video card and then exhausted out the back Um this is going to give a pass through that blows air up through the through the through the the edge of the video card away from the connectors away from the the mounting plate over the CPU over the memory over the voltage regulators all of that stuff And then out the back I think I'm really interested to see how this works I hope it's quiet That's my big take away
Brad: [01:14:37] they claim like you know like like we said before they claim that it is both quieter and more efficient than their old design which you said was bad So probably wouldn't be hard to do better
Will: [01:14:47] I mean look it's not bad It dumps a shitload of heat out It's just really loud
Brad: [01:14:50] It's is very loud Yeah Yeah Um
Will: [01:14:54] so the small board with in necessitated uh power been redesigned So instead of using the standard eight pin you know two four pin connectors
Brad: [01:15:03]uh is that why they’ve gone to this new kind Okay I see
Will: [01:15:05] Yeah So uh the new connectors in the middle of the board it's a 12 pin connector all the boards that so the Nvidia design reference boards have the 12 pin connector Some of the partner boards will not it's up to the partners whether they want to do that Some of the partner boards don't use the full blow through cooler they're using their own cooler So um yeah and of course there would be partner boards that are like water-cooled with mounting blocks and all that stuff as well Um The big thing about this generation in terms of partner boards is that the 3090 what would have been the titan before This is why they changed the name I think is that the Titan was only manufactured by Nvidia Uh the 30 nineties are manufactured by partners as well So those $1,500 boards are available across the board which means we'll also see some like $1,500 We'll see some of those 3090 boards that have like weird memory clocks and memory configurations and stuff like that presumably Um I think this I think this Like I'm I'm stoked about the cooler design It looks cool Uh so the the the fan that's furthest away from the connector edges pulls air through the fins of the cooler and blows it up toward the CPU and the memory the fan that's closest to the mounting edge of the card sucks air in and then blows it out the back So there's still an exhaust component Um but it's also it also is like facilitating air flow through your system
Brad: [01:16:25] It's basically it's it's exhausting air out of two different spots on the card the top the top front and also the rear out the back of the case which probably I would assume is better for cooling But one thing that um the one question is raised for me is whether there's any concern around you know that front fan like you said is exhausting right over the motherboard components and into the CPU fan Like is there any concern around heating that air before you send it over all those other components Like is that going to impact CPU cooling
Will: [01:16:50] I mean the ambient inside your case is going to be hotter than ambient in your in your house So yeah there's going to be an impact but will it matter No I don't think so Um and and if you are concerned about that what you should do is is do a positive pressure cooling setup where you have an intake fan like most most cases for enthusiasts Now have the option of doing fans on the top above the CPU and the voltage regulators and all that stuff So put a 12 or 14 centimeter fan and they're blowing in and that'll give you a nice Versus a cooler there which I don't think this will be a problem Uh somebody asked if the video cards would the way the vape sometimes vapor chambers and he pipes can be set up in a way that they don't work in certain orientations It's less of a problem now Cause there's like wicking materials that pull the the condensed liquid backup um on the inside of the pipe sometimes they said it wasn't they said it'll work in any orientation So if you want to vertically Mount your card either horizontally across the front of your case window or vertically up and down you can do that Um the the let's see before we move on the 3090 is a big chonker of a card
Brad: [01:17:58] it's a three slot card
Will: [01:18:00] three slots It's like 316 millimeters or something from edge from short edge to long edge Um
Brad: [01:18:07] I actually looked up the specs on my case just out of curiosity it does technically fit in my case but you have to remove like all the front hard drive racks
Will: [01:18:14] Oh wow
Brad: [01:18:15] to make extra space for it Um
Will: [01:18:17] it will fit in my case without having to remove anything, hint hint.
Brad: [01:18:21] I've got room to just saying um the other thing I was gonna say is that that connector being moved back more and more toward the mounting bracket I mean it sounds like about it's about halfway down the card
Will: [01:18:31] the 12 pin?
Brad: [01:18:32] Yeah Um
Will: [01:18:33] Yeah the 12 pin almost almost above the GPU It looks like which is
Brad: [01:18:36] I'm sure that was primarily because of the shorter PCB but but it does make me wonder if that is also a concession to mounting in more cramped cases because if you they're there they're probably a lot of cases out there where the card itself would fit with whatever kind of drive mounting or whatever is in the front of the case But once you add that extra horizontal space of trying to connect this stick this big plug in there it might no longer work So like I've seen people complaining that they moved the the um The the power connector on the card back
Will: [01:19:06] From the end, Yeah
Brad: [01:19:06] But it seems that seems useful for people that have uh smaller cases that that actually will probably give you some more clearance
Will: [01:19:12] I'm trying to think I think so I think my 2080TI has the edges on the top edge of the card like away from the edge furthest from the PCI express connector So it doesn't stick out the it doesn't stick out the long end It sticks out the top but you're right My 1080 definitely had connectors on the end
Brad: [01:19:30] I still have a 1080 and it is in yes it is It is It gets a little cramped with that That's like an extra inch of space or something almost with that you know those tie those kind of inflexible cables sticking out the side So hopefully this will be easier for more people to fit into their cases with that connector moved
Will: [01:19:48] Yeah Well and the 3070 and the 3080 are much more reasonable sized cards
Brad: [01:19:51] Yes Yes Those seem pretty standard
Will: [01:19:54] Um so I think let's see the last thing is the they're adding um They're adding where are we There we go They're adding more stuff to the they're They're this Nvidia broadcast tool um is the successor to the RTX audio uh beta that they released this spring I think they rolled that
Brad: [01:20:15] there They're rolling the RTX voice stuff into this broadcast app Right
Will: [01:20:18] Yes It seems like it um
Brad: [01:20:20] The thing seems pretty cool
Will: [01:20:22] I like as somebody who's streams I think in a world where we weren't all stuck at home doing a lot of zoom calls this would be a novelty for people who stream but in a world where you have zoom calls this sounds like magic
Brad: [01:20:35] Yes you know they showed on that stream The some examples of the dynamic background removement removal stuff And it seems like it works pretty well from from what I saw
Will: [01:20:42] You mean in there controlled situations that worked extraordinarily well I'm very interested what it looks like in people's houses with their shitty webcams and their bad lighting Um but yeah so it uses machine learning They trained on people and chairs so it removes everything but the chair and the person um it seemed to do a shockingly good job getting rid of the like Would that gap between hair where there's some transparency where the edges of your hair are poking through which is the hard part for green screens and all the other depth camera vision cut outs I've seen
Brad: [01:21:14] my guess is for uh for these small inset type of framing that a lot of people use for streaming you know where their their webcam feed is quite small in the overall frame It's probably gonna be totally fine even even a little artifacting like that probably won't be noticeable at the size that most people are gonna be using it at
Will: [01:21:30] I'm really interested to see what it looks like when you put a good camera on it So
Brad: [01:21:33] I am I so I I can't make use of this right now cause I'm not on an RTX card but actually they offered us early access to this to this application But I can I I can't try it but I think we talked about on the list Was it the last patron we talked about Yeah we talked about all the stuff I've got coming for work I've got a I've got a nice high end mirrorless camera on the way So provided I provided I get an RTX card in here Like I'm really fascinated on how super nice looking camera feed like that is going to play with this It's probably worth mentioning how they actually integrate this into your streaming setup which is the big thing I was a little concerned about here
Will: [01:22:07] Yeah you don't have to stream through GeForce experience to make work, that the important part.
Brad: [01:22:10] I was yes So I feel like maybe a slightly shittier company you might have tried to go that route of like You know this is trying to capture a new market of like Hey we can like sort of force people into our streaming pipeline by giving them these cool features but they
Will: [01:22:23] If you want people to use this stuff that are professionals then you have to make it accessible to all to whatever tool chain they currently use
Brad: [01:22:28] Right right And that that's what they've done here So I maybe I was worrying for nothing but uh that is how this works That takes in your um takes in your webcam feed your microphone like your your game audio or whatever other audio sources And it sounds like it's going to spit them out as virtual devices in windows
Will: [01:22:45] Yeah it seemed like it seemed like with the there's a newer version of RTX audio coming with this as well That will let you like assign it per input So like if you have an input for your discord chat then you can you can just do RTX audio on the discord chat If you have a good mic stuff like that
Brad: [01:23:02] Yeah So it almost sounds like to me correct me if I'm wrong it sounds like this kind of works the way that voice meter does where it's creating virtual audio devices as outputs in your system And then you can just you can assign those out in different applications
Will: [01:23:13] That seems right And they said that it will work like running compressors like running normal audio tool chain stuff on it Won't Jack up the impact for in let's see what he said He said if you're like I think you said reasonable tool chain So I assume that means like noise escaped and compression not like reverb and stuff like that Um
Brad: [01:23:36] I'm surprised they were fairly candid about the performance impact of the stuff which is bigger than I expected
Will: [01:23:42] Yeah
Brad: [01:23:42] so they they mentioned like the RTX voice beta they were looking at an 8.5% hit to your frame rate If you had it enabled they've got that down to yeah they've got that down to 3.6% in this release version which is not that bad but they say that the go ahead
Will: [01:23:56] was gonna say presumably that's on stuff that uses the RTX RT cores and the and the and stuff like that So I would think if you're running something CPU limited you'll see a smaller head
Brad: [01:24:06] maybe so
Will: [01:24:06] So if you're running like 300 frames in Overwatch you're probably fine
Brad: [01:24:09] right Um
Will: [01:24:10] running control you're going to see a performance hit.
Brad: [01:24:12] It did make me think like if you remember back when that RTX voice beta first rolled out there were people like instantly like the same day or like Hey you can just like hack the hack the installer very easily and make this work on a 1080 or whatever on the Pascal cards But seeing this performance hit made me realize like doing that kind of like you were saying with running Ray tracing on the shader cores in a 1080 probably was working that card pretty hard to do the stuff right
Will: [01:24:36] I would think so Yeah
Brad: [01:24:37] you were actually pumping a fair amount of juice into your card to make this work
Will: [01:24:42] guess is that on a non RTX card you're probably running math on shader cores That's designed for it's dedicated units on an AR on a Ray tracing card And as a result you're on a dual PC streaming setup You're probably fine on a single PC streaming set up your game Performance is probably going to take an enormous hit
Brad: [01:25:00] That's the point I was trying to make is like using that on a on a GTX card Uh would be prohibitive if you're also trying to crunch a game at the same time Um and the the background removal on the webcam feed they're saying is that at the moment is a 15% hit to your frame rate which is not negligible
Will: [01:25:17] but but they also made it abundantly clear that they feel like there's a lot of room for improvement there.
Brad: [01:25:21] Yeah Yeah They they they say they're going to keep working on reducing that But as of right now that's like depends on the game Like some games that wouldn't be a problem Cause like you said you're running games like Overwatch it and sane frame rates already but there are certainly games out there where I would not want to sacrifice 15% of the performance for a feature like that
Will: [01:25:39] my takeaway on this both of these features is that they're probably going to get RTX voice down low enough that everybody is going to want to run it Uh if you're playing games online with friends so that they don't hear your keyboard and stuff like that uh if you are especially if you're doing a lot of voice calls it's worth putting on cause like I can run the fan in my office and nobody hears when I'm on a zoom call
Brad: [01:26:03] Like you said if you're not playing a game at the same time it doesn't really matter how much it's working The graphics card
Will: [01:26:07] Exactly Um when my guess is that the broadcast the greenscreen cleanup thing you'll probably only use if you have a two PC setup or if you have a like a 3090 if you're like I my guess is that's only something that streamers who are fairly pro like pro like professional or pro you know Pro-Am and have multiple PCs doing their setup We'll do
Brad: [01:26:30] Of course those people probably have green screens and other solutions for that stuff already
Will: [01:26:34] I'm going to go and tell you the green screen It's not great Brad it's a pain in the ass
Brad: [01:26:38] but it's right there I see it right behind you How bad
Will: [01:26:40] That’s the problem You can see it Right But no the um the problem with green screens is that in order for them to work well and look good and give you a clean no bounce back result where your hair and like your skin tone doesn't get pick up a little bit of the green You either have to have a lot of distance where you have to have everybody involved lit really well And it's hard to do that Unless you have a lot of space like like for example when we were in the whiskey basement we used to use the green screen down there a lot norm and I would stand 10 or 12 feet off of the green screen to make lighting it easier And and would this in most people's houses don't have that kind of space So so having like the green screen is not a fantastic solution That's all I'm going to say
Brad: [01:27:20] Is this the type of thing Is this kind of keying the kind of thing where if cameras were shipping with dedicated depth sensing hardware or you could even maybe buy an external like USB depth sensor that you could pair with a camera like that that this problem could just be solved more easily That way
Will: [01:27:35] As somebody who has two USB depth cameras connected to their two PCs on their desk Right now I will tell you that they probably provide valuable data but the the the depth camera resolution is blobby enough that things like your hair end up with with outlines around them of the of the real background So probably not
Brad: [01:27:57] Okay Right They're not there yet on that stuff
Will: [01:28:00] Um and I
Brad: [01:28:00] May when maybe when maybe when the LIDAR is out there maybe you get a LIDAR sensor for your
Will: [01:28:04] the lidars were we had LIDAR that's the connect one It was great We love that thing
Brad: [01:28:07] Oh good Okay
Will: [01:28:08] Um I feel like you get a little bit cleaner data on the newer depth sensors that are on the iPhone Like the newest sensor I have here is probably two years old so they've they have progressed dramatically in that in that time But right now my feeling on depth sensors is that they're really really good for things like uh collision avoidance on robots and drones and maybe not great for fine detail work
Brad: [01:28:31] Okay That makes sense
Will: [01:28:32] Um I think that's it
Brad: [01:28:34] Okay That's a lot
Will: [01:28:35] I think we're done We've won
Brad: [01:28:36] that's a lot man
Will: [01:28:38] Congratulations
Brad: [01:28:39] you're you're not wrong for as much as I was saying these cards are just sort of the performance bump to the previous cards like around the edges There are a ton of new features here It's just not there are no new graphics features on the order of RTX You know like that that was like a RTX was like agenda that's that's a generational change
Will: [01:28:56] That’s an every 10 years
Brad: [01:28:57] Um but but with you know all the direct storage stuff and the Kind of the addition of DLSS all over the place And uh you know looking down this list there's a lot of new stuff that you can do with these things that doesn't necessarily pertain directly to graphics
Will: [01:29:13] I um so when I said it wasn't that what I meant specifically is previously Like usually when you think of the tick as a new architecture and the talk as a revision of that new architecture this is a revision of a new feature but it is a new like ampure here represents a new
Brad: [01:29:34] It’s a new micro architecture
Will: [01:29:35] A new spin of the Silicon It's more than just a new spin of the Silicon I guess is what I’m saying
Brad: [01:29:39] Yes You're right Wait did you actually just lay out tic talk in the sensible way Cause I'm pretty sure there were a lot of people out there who claiming it was supposed to be the opposite back then A lot of people like I remember people saying that the talk was supposed to be the the big news the big new change because in their minds talk is louder than tick
Will: [01:29:57] I think Intel was the proponent of that theory
Brad: [01:30:01]That’s insane to me Tick literally comes before Talk talk is the second of two
Will: [01:30:06] I had a conversation with an Intel person when they rolled this whole thing I was like look man tick is the first one And I don't care if talk is louder tick is the first And then the talk is when you make the new thing Good And then the tick is the new thing And then the talk is when you make the new thing they've ditched that they decided it was confusing and they got rid of it So
Brad: [01:30:26] it’s not talk tick Okay
Will: [01:30:28] Look that would be a weird clock I don't want to talk tick clock
Brad: [01:30:31] Wait what if we can make time run backwards
Will: [01:30:33] No no I don't want to go No I don't want to do the last six months again
Brad: [01:30:37] that's a good point Actually We can be careful what I wish for
Will: [01:30:39] Maybe forward It's faster would be bad anyway Brad, I think this is our one year anniversary of the launch of the show
Brad: [01:30:46] Is it really like to the
Will: [01:30:47] so Well
Brad: [01:30:48] pretty close to the episode
Will: [01:30:49] in that where you get to that Yes The thing that you just said um We cause we skipped one episode As I recall this in this in this the first year um
Brad: [01:31:01] It's kind of flew by this wild I don't know Where what did we talk about
Will: [01:31:05] the months went really fast and the days were interminable Uh but it's time to thank our patrons
Brad: [01:31:12] Thank you for getting for free for getting us here and keeping us going
Will: [01:31:17] Uh if you would like to find out how you can support the podcast you can go to patrion.com/techpod And for as little as $2 a month you can get access to the fabulous tech pod discord There's a new channel We didn't talk about it last week
Brad: [01:31:31] Oh yeah Oh, hang on. I'm struggling here. Don't tell me I don't. What is the new channel
Will: [01:31:41] Wait, isn't new ass computers, the new channel.
Brad: [01:31:44] that's been around for a little bit? The name may have gone through a couple of permutations.
Will: [01:31:49] I feel pretty good about it. I like where we're at
Brad: [01:31:51] I don't know. I think that they, that that name may have shelf life. I think theres...
Will: [01:31:56] Oh, I think every name has a shelf
Brad: [01:31:58] think, I think there's a better joke to be had there.
Will: [01:32:00] I'm totally down with a better joke. Um, but yeah, uh, the tech pod discord is full of brilliant, wonderful people who are talking about all of the things that we talk about at a much deeper detail level.
Brad: [01:32:12] There, there are people on that discord who absolutely would have sat through this presentation and fully grasped everything that was being discussed.
Will: [01:32:17] And like,
Brad: [01:32:18] There are definitely, there are definitely some graphics programmers on that discord.
Will: [01:32:21] Yes, that is 100% true. That they're the ones that were screaming at their radio through the entirety of this episode. Um, but, uh, if like from. Stuff like the Mister and I got a ton of help with my home assistant set up last week. Uh, it is an unbelievable resource for learning more about how the world around you works, which is what I love and, uh, people.
Brad: [01:32:44] Do you know what I saw somebody do on there?
Will: [01:32:46] Ooh.
Brad: [01:32:46] Um, I've talked about how I have my Astro Mix-Amp mounted under my desk, and I use some of that 3M dual grip that we talked about a little while back, somebody, somebody did it with that, the nano bubble stuff.
Will: [01:32:57] Oh, did it stick?
Brad: [01:32:58] Yeah. They just posted a photo that they've got the full on mix amp stuck to the bottom of their desk.
But there's no adhesive. I mean, I know there is on the backside where it's stuck to the desk. Right. But the isn't the, isn't the actual nano bubble material. Isn't it it's literally just like suction or refrigerator or whatever, holding it on their right note. There's no glue. There's no glue. Oh, that's wild.
I might have to get some of that stuff.
Will: [01:33:18] You should. It's amazing. It's like 15 bucks for a sheet that will last you a year. If you, if you cut it to small places.
Brad: [01:33:24] I love to mount .
Will: [01:33:26] It is. Hold on, phrasing there.
Brad: [01:33:29] I love to Mount things. Is that better?
Will: [01:33:31] Nailed it much better.
Brad: [01:33:33] That totally makes the difference. Um, I have with this episodes long already, I shouldn't just sit here and ask questions, but do you think something, do you think something like that, and this is a ways off, cause I still have this crappy old desk, but it was doing what I needed to do, but one day I will have a different desk. And I'm pretty sure I'm going to want to Mount like surge protectors and my router. And unless I have, unless I lived the true dream and have a rack in my basement and things like the ratter live down there. But,
Will: [01:33:57] the basement rack,
Brad: [01:33:58] but if not,
Will: [01:33:59] Brad, I looked at racks the other day.
Brad: [01:34:00] they're expensive.
Will: [01:34:02] You can get a half like a one that would fit under your desk. Cause it's not that bad.
Brad: [01:34:05] who wants a half rack
Will: [01:34:07] I kind of want a half rack I don't, I don't want to have to move a full rack. They're really heavy.
Brad: [01:34:11] if I buy a rack, I want it to be as tall as me, but anyway, For the moment, let's say I wanted to like mount my router and network switch and a couple of surge protectors underneath the desk, like literally upside down. Do you think something like the nano bubble would be sufficient for that?
Or is that asking for it?
Will: [01:34:26] Cause the cables on the power,
Brad: [01:34:28] Oh, too much tension
Will: [01:34:29] it's too much tension on the cables. Um, I just screwed some of those baskets into the bottom of my desk and then I threw all that shit in
Brad: [01:34:35] just, just lay it in there
Will: [01:34:35] can't see it unless I'm laying on the ground. So yeah.
Brad: [01:34:39] Okay. Good enough.
Will: [01:34:40] Um, I would think about putting some under lights on the desk though.
Brad: [01:34:43] Ooh,
Will: [01:34:43] downlight the desks, get a little bit of a purple.
Brad: [01:34:46] I've got a backlight and I think I've, we've talked about it. Just got the, the, the led colored led strip on the back of my desk. I can make it orange or purple on the wall.
Will: [01:34:56] the two of the only two true colors,
Brad: [01:34:58] Those are my two favorites.
Will: [01:35:00] um, uh, it's the, it's the time. So yeah, if you're curious about this, you want to check it out. The address again is patrion.com/techpod. Uh, as always, if you can't support the show, that's fine. We're we're, we're doing great , thank you all for your support
Brad: [01:35:11] Thanks. Thanks for just thanks for
Will: [01:35:12] your
Brad: [01:35:12] You can tell your friends.
Will: [01:35:14] I love it. I love seeing people on Twitter sharing, word of, of the things that go on in the tech pod.
Uh, we really love that. It's been great. Um, and thank you as always to our executive producer level patrons, Andrew Cotton, David Allen and Jacob chapel. Thank you all so much
Brad: [01:35:30] Yes. Thank you everyone.
Will: [01:35:32] Uh, and we will be back next week with more Brad and Will made a techpod. I think I've been fooling with Moonlight, Brad.
Brad: [01:35:40] Oh, no, no.
Will: [01:35:42] Yeah. I was able to play Epic store games on my steam link in the living room.
Brad: [01:35:46] Wait, how are you, how are you running Moonlight on a steam link or are you just capturing the, okay.
Will: [01:35:51] Moonlight runs on steam link , dude,
Brad: [01:35:53]i didn't know that.
Will: [01:35:54] I can run up to like a hundred megabits per second. The video stream of my wired network
Brad: [01:35:58] How does it, is it better than the steam home streaming?
Will: [01:36:01] It looked better. It seems a little less jittery.
Brad: [01:36:03] Interesting. How's the latency cause
Will: [01:36:07] was able, I was able to land tricks in Tony Hawk.
Brad: [01:36:10] Okay. Doesn't Moonlight use the nvidia streaming tech.
Will: [01:36:13] So you basically, you, it uses this stuff. Well, you should talk about this in episode.
Brad: [01:36:17] Okay. Yeah, there's
Will: [01:36:18] This was a tease.
Brad: [01:36:20] Yes, there are. Let's
Will: [01:36:21] I want to give it all
Brad: [01:36:22] let's say I forgot. I'll talk about that. Let's say there are quite a few options for both in home and across the internet streaming of your games. And maybe we'll maybe we won't, maybe we'll go through all of them at some point.
Will: [01:36:34] Dude, we should set up that Moonlight internet node thing. I should set that Moonlight internet node thing and see if you can connect to it
Brad: [01:36:40] Oh Weird.
Will: [01:36:40] and see how they'll see if you can play from my house.
Brad: [01:36:44] I didn't know that existed. That, that
Will: [01:36:47] this is, this sounds like an episode.
Brad: [01:36:49] could be useful for my day job, actually.
Will: [01:36:51] Oh yeah. I didn't even think about that. Yeah. Couch games. I, what I want to know, here's a question for the audience before we get into this, is there a way I can plug. A USB connector from my computer into the Mister and then pipe, the video output of the Mister through some magic.
That's going to let like you and I sit down and play, say a two players SNES game running on the mister In my, on my desk, across the internet.
Brad: [01:37:20] Well, if you have capture hardware, you can just capture the video
Will: [01:37:22] Yeah. But how do you get the input into the mister
Brad: [01:37:24] Yeah. But the input is the question, I guess that would probably just have to be coded into the mister System. Like theor cores, something would have to recognize that input.
Will: [01:37:32] I'm asking. Cause I, I want the people I want, I bet somebody knows. I bet you can just parsec or something.
Brad: [01:37:37] I'm trying to think parsec does have, or we can at least tie into some kind of gamepad emulation format,
Will: [01:37:44] I might have to, you might have to use like an Arduino or something in between. So it sends, so the Arduino would show up as a game controller and then for the, for the mister But it's programmed by the PC and it's just passing signals.
Brad: [01:37:57] we are totally just cannibalizing material here at this point, but there is, there is a support page I believe on parcel on the parsec site, uh, that tells you how to play Xbox one games remotely. there is a way to, there is a way to emulate the input of the second player on an Xbox one from a PC through Parsac.
And I'm not sure. How it works, but if that's happening at the D input or expert input level, you might actually just be able to map that to,
Will: [01:38:21] to the minister.
Brad: [01:38:21] a virtual game pad that would work with the Mister that's possible. You should look into that.
Will: [01:38:25] I will look into that. This sounds like a fascinating episode perhaps for next week.
Brad: [01:38:29] test that out and report back.
Will: [01:38:30] Yeah. Uh, always a pleasure, mr shoemaker.
Brad: [01:38:33] Yeah, thank you. I know this was a dense episode. I hope that it's a, I hope it was illuminating for people though.
Will: [01:38:38] I'm going to go take a nap now.
Brad: [01:38:40] That sounds good. It's very hot. Can, I don't know if I've, I haven't checked the air quality today. I don't know if it's safe to open the window or not. Let's see.
Will: [01:38:47] we've had windows open
Brad: [01:38:48] It's pushing it's
Will: [01:38:49] I have opened last night. It was great.
Brad: [01:38:52] okay. We're at we're at 8:00 PM.
2.5 of 77. What do you think?
Will: [01:38:56] Oh, that's, that's fucking fantastic. Compared to what we've had lately. Hold on. Let me,
Brad: [01:39:00] Yeah, but would you leave the window open all day with that?
Will: [01:39:02] 69 75.
Brad: [01:39:06] but stolen the yellow?
Will: [01:39:07] That's in the, yeah, that's in the, Hmm.
Brad: [01:39:11] approaching the danger zone for me.
Will: [01:39:13] Yeah. Like once we get into the yellows, when I turn it down, let me see where we're at. According to my home assistant install, we're at 61.
Brad: [01:39:23] Oh, that's not bad. That's almost in the green.
Will: [01:39:26] Yeah. It's pretty, pretty good.
And right now outside, it is 68. Oh, hold on in my backyard. It is 79 degrees right now. Holy shit. It's already 79 degrees. It's only fuck. It's only 11. O'clock what the hell? I'm going to go edit this before it gets 90 in here. See all next week. Bye everybody.