John Dellabona (4:53)

That's fantastic. I mean, that was one of the features and I'm glad that you brought it up because I have kind of the list of the big changes with Blackwell. Like, give me an example of, you know, a workflow that would take advantage of kind of the neural network shaders that are going to be released in Blackwell. Is that specifically like a simulation thing? Is that a media and entertainment kind of pipeline workflow? Who will take advantage of that and what, what verticals will take advantage of the, the new neural network shader technology in Blackwell?

Logan Lawler (5:20)

Yeah, the, you know, the exciting thing is developers are going to continue to adopt and work with this technology to integrate it into more applications. One example that I'd love to talk about is DLSS4, which is integrated into D5 render, for example, which is Used for things like architectural visualization. So DLSS4 for the first time is able to almost see into the future by generating multiple frames based on what's in what it's interpreting and seeing in real time on your screen. So if you're, you know, if you're in a real time engine and you're developing say a building design, it can, it can generate additional frames with AI, thus massively boosting your performance and giving you that real time interactivity. And because AI is being able to generate these scenes and offload that onto the tensor cores, as a result, you can work with scenes that are much more photoreal than ever before, because your traditional CUDA cores in concert with the tensor cores are working much better together and there's more headroom to really drive through photorealism. So DLSS4 is one example. We're excited to see these neural shaders continue to get adopted across different applications.

John Dellabona (6:50)

That's really cool. I mean, I love that, you know, bad probably analogy, but you're kind of drinking your own Kool Aid, bringing kind of AI into, you know, the actual GPUs, which I, you know, I absolutely love. Like in terms of that, you know, I gotta definitely have a question. I think that it's very interesting to hear kind of the frame prediction, you know, and I correct me if I'm wrong on the number, but I think it was like, it's like estimating like three pixels ahead or so. I think that's the number I heard and that could be wrong, but it's with the pixel estimation ahead, like in a traditional pipeline, it's kind of allowing you to kind of see into the future. Like what are some of the specific ISV applications that this is going to be extremely relevant to?

Logan Lawler (7:34)

Yeah, I think there's quite a few. So for one, virtual production is a big one. When you're on these big film sets, we're now at a point where actors are in front of these giant display walls and the 3D world is being projected right behind them and they're basically able to be immersed in the film. But that requires real time graphics and that requires, in the case of DLSS4, real, real time AI, right? To be able to drive those frames and create that huge, huge scene with complex geometry and you know, predict those frames into the future to make that hyper detailed scene all work in real time behind the actors. So I think virtual production is one great use case. Another that I previously mentioned is architectural visualization, which requires highly advanced techniques, whether it's simulation or Photorealistic ray tracing to do accurate light studies. All of these are computationally very expensive. And so having AI be able to step in and help and drive those real time graphics and improve the performance is going to be, is going to be very beneficial. So those are two example industry use cases I can think of.

John Dellabona (8:49)

That's awesome. So there's another big thing, another big announcement with blackbill, which is four two two. Can we go into a little bit of that? Because I know, for example, those that are in media entertainment that specifically work on video are going to be very excited to hear about this.

Logan Lawler (9:05)

Yeah, absolutely. So long requested feature for Four2Two support and we're happy to announce that with this generation of Blackwell for RTX Pro, we're supporting Four2Two encode and decode for the first time. And that's going to be a large quality jump for creators that need higher color accuracy when they're, you know, editing creative footage, doing things like color grading. But then on the encode out process, it's also going to mean that you're spending much less time waiting and actually able to enjoy the encoding process, if you will, while your files get exported out lightning fast. You're not going to be sitting and waiting like you used to with a 4,2 2 format. It used to be a CPU bound workflow. And now being able to put this onto the gpu, it means that your, your videos and your encode encoding out of 422 files is going to be very fast as well as the real time usage of that footage. You're going to be able to, you know, edit and color grade with those kinds of things much more fluidly and interactively. So we're really excited about the GPU support there.

John Dellabona (10:17)

Yeah, I know that and I mean we are Dell technologies at the end of the day. But I mean I think that is one thing that Mac users have kind of had and why they've gravitated more towards Mac for use and you know, whether it's Photoshop, other different things. And this really, you know, helps to bring to bear, you know, a lot of great things that you can do with PCs as well as GPUs. But now kind of being able to connect that and take out a huge pain point is amazing. It was absolutely amazing. So I know I can't remember you specifically, you support mobiles, correct? Or do you do the fixed GPU side of the house? Well, today you're doing everything, but for, for the listeners you support, I think it's mobiles, correct?

Logan Lawler (11:00)

Yeah. So my Current focus is mobiles. I've been lucky to work on a lot of different products at Nvidia, so I'm comfortable kind of across all our solutions. I've worked on desktop products, data center products, but my current focus is mobiles.

John Dellabona (11:13)

Is mobiles. Okay, well let's kind of start with mobiles and then we'll, we'll kind of move to fix. So you know we kind of talked about you know, the really three big kind of changes. You know, functions feature now as feature set. But let's kind of get into the Nvidia, the overall kind of GPU stack it looks like and correct me if I'm wrong, but kind of in the current kind of, you know, ADA lineup, kind of running from a 500 up to a 5,000 kind of remains the same in terms of the naming nomenclature. But there's a lot kind of changing under the hood. So you know, we could go through every GPU which we don't have that much time and we're not going to do. So we'll just kind of start at the top end of the stack. Right, let's start with the, you know the Blackwell 5000 is kind of give me the high level on increase in performance kind of gen over gen with with against ADA and kind of in terms of CUDA Tensor ray tracing kind of the big feature, you know, performance highlights, you know, the increase in vram, all of that kind of good stuff.

Logan Lawler (12:16)

Yeah, of course. So we are launching Blackwell across across the full laptop family stack. So that's the RTX Pro 5000 Blackwell generation all way through to the RTX Pro 500 Blackwell generation. So in total we've got six GPUs we're announcing there for the, for the family platform. At the high end the 5,000. Something we're really excited about is we bumped up the memory at that class level to 24 gigabytes of ECC memory. Previous gen we capped out at 16 gigs of GPU memory there. So having that but significant bump from 16 to 24, you know, our professional audience is going to be able to work with larger 3D models, larger AI data sets, work with bigger multi app workflows and do that all on one gpu. The other thing I'd like to highlight, you were talking about cores and things like that. So some of the technologies we were talking about earlier, all that goodness of Blackwell funnels through into all of these GPUs. So that's the new ray tracing cores. For faster ray tracing we've got the new Tensor Cores for processing AI, we've got the streaming multiprocessors with the neural shaders. We've got new video encode and decode engines with 4, 22 support. So all those great architectural features are funneling through to the products. And you know, performance, it always varies by application. But I think if you look conceptually even at like the ray tracing cores, so those deliver up to two times the throughput for ray tracing. And then if you look at things like AI, FP4 support effectively doubles your, doubles your performance while having your memory requirements like working with AI models. So there's a lot of exciting performance there. And you know, there's, we're, we're going to make benchmarks available for the product so customers can kind of check out per application what kind of exciting boosts they can get to their, to their application workflows.

John Dellabona (14:19)

I mean, that's. Yeah, that's fantastic. I'm looking forward to seeing the data. I do have a question. I think this might be. And it's not a head fake. That's probably the wrong, the wrong terminology. But you know, when I first kind of saw the kind of the let's just use mobiles, right. Kind of the lineup, you know, I'm looking at it kind of now, you know, we're going through cores, we're going through and I got to kind of AI tops. And previously I'm, I mean working with Nvidia, working with, you know, Dell Promax workstations. I'm very much used to seeing, you know, kind of spec sheets around tflops. So I think that might be a bit of a head fake. Cause when I saw the number I'm like, oh, this is new. I've never seen this before. Maybe if you could explain a little bit about kind of AI tops and why you're using that as kind of the comparative benchmark now in Blackwell and kind of the change from TFLOPS.

Logan Lawler (15:13)

Yeah, so our, our AI tops performance numbers, they're, they're actually using the, the FP4 data format that we're supporting. So we, number one, we like to highlight the new features and be able to talk about the, the theoretical AI throughput and the max possible AI performance that you can get out of a gpu. And two, so tops are basically tera operations per second and they're kind of an industry standard metric that a lot of different companies and enterprises have been referencing to kind of look at what's the theoretical AI throughput of, you know, a specific product or gpu. So we Think it's important to provide the AI tops number out there so customers can get a sense of purely from the gpu. You're getting up to, you know, a thousand eight hundred tops from a single processor. It's immense. And I mean you look at CPUs and NPUs and those things being able to provide, you know, sub 50 tops and around that under 100.

John Dellabona (16:26)

Yep.

Logan Lawler (16:26)

Yeah, right. It's, it's really important for customers to understand that when they're buying a workstation, these GPUs are going to be really critical to driving their, their AI workflows. And so the TOPS number is just another thing we provide to support that and really illustrate that a large amount of this AI compute is being driven by the gpu.

John Dellabona (16:48)

Yeah, I mean it makes sense. I mean AI tops that number, you know, with whatever it was, Meteor Lake and Intel kind of has been a number that we looked at. So it's much better than Logan Lawler doing his. And I'll tell you a funny story is that someone asked me that question about, I don't know, I mean I was enrolled for about five months, so this would have been like last summer. And someone asked me, hey, how does an NPU compare to a gpu? And I, I mean I was burning up the Internet like trying to figure out the do, do, do do, trying to find the calculation, right. And it's easier just to provide it because that's kind of the number. So it ma. It makes total sense. I mean, you know, I, and I love the, the comparative because you're right, but I want to kind of take a step from the top of the stack to the bottom of the stack. And you said, kind of an interesting point is, and something that I really want to drive home for people is that, you know, when it comes to, I mean Nvidia is kind of the gold standard, we all know that. But as kind of AI evolves, there's really, you know, you think the ADA 5000 on the mobile side or not ADA, the Blackwell 5000, you know, you're doing your training, you're doing high end inference, you know, all these things. But even at the 500, it. You can't forget that there's still going to be applications. If you're doing a refresh cycle of three, five years, whatever it is, is that AI is going to look very different today than it does five years from now. And more and more things are going to start to be run locally, you know, in the. Where do you think the sweet spot? If you, I Mean, obviously applications are different, but you know, as you kind of see the, you know, everything evolve the requirements of offloading to the gpu, where do you think kind of the entry or sweet spot would be for someone who's like, hey, I want to future proof this workstation, like I need it for X. But I'm not, you know, I'm not running simulation, I'm not doing crazy reality capture, I'm not doing any of these crazy workflows. But I just want to future proof my system. Do you think you, is it of the 500, is it the thousand? Where do you think the sweet spot would be?

Logan Lawler (18:49)

Yeah, I mean it's a tough question. I know it definitely, it definitely depends on what the customer expects and wants to do. And then another thing to keep in mind is, you know, even at the 500 class, you know, although it's capable of many things, anything it's capable of is going to scale up drastically in performance as you go up the stack of products. So you could do generative AI on an RTX Pro 500 and generate new images from text prompts, for example, and that would be significantly faster than doing it on a CPU or an NPU or integrated graphics. And then, you know, and that could be, for example, let's just say hypothetically that's an hours to minute story. And then you know, on the 5000 class it's A, it's a minutes to seconds story. Right. So it depends on performance requirements. But additionally I would say there is some workflow segmentation. So for the RTX Pro 500, I think one thing that I see is AI is being infused into everything. And it's not just traditional 3D applications and AI development pipelines. It's in, it's in the things you and I are doing right now, right? Video conferencing. There's suites of AI features being adopted into video conferencing to improve quality of your voice and your visuals or keep your eye days tracked on screen if that's something you want to do. Those things are all compute expensive and can be offloaded onto the gpu. So even for like an everyday knowledge worker, an RTX Pro 500 is really valuable and we see it too. And even like video streaming, if you're watching YouTube videos and they're low resolution, on an RTX GPU you can actually use AI in real time to, to up res that low resolution footage you're watching to match the quality of your high resolution display. So these are kind of like everyday features that I expect are going to continue to grow and I think are going to make it more and more necessary to have even, you know, an entry GPU like our RTX Pro 500 available in your laptop. And, and then, you know, of course if we go all the way up to, you know, and the RTX Pro 500 too can also do other things like 3D rendering of smaller models and you know, using AI features in Photoshop and video editing applications. So it's great for that too. But I think when you scale all the way up to the 5000 class, then we are talking about completely different classes of workloads, right? Data science, AI development, large scale VR running simulations, highly detailed ray trace rendering. So it's not like a one size fits all for every gpu, but you know, we have a GPU for everyone, so.

John Dellabona (21:51)

Exactly. And it's a good point. It is like an, and hours to minutes and a minutes to seconds as you kind of go up the stack. But you, you, you went exactly where I was hoping you would go. Is that like there are so many, and I'm not going to say they're under the COVID type workflows, but there's AI things that are running that people, it's not that they're bad, but you're right, this background blur, things like that people don't know, is necessarily offloading, you know, to the gpu. So, you know, my recommendation is that even though you think you might not need a gpu, if you're going to keep this thing for a while, you're probably going to want a gpu. And even if it's in the smaller end of the stack, like it still makes sense because would you rather spend hours doing something or would you rather spend minutes or seconds? So let's get into, you know, we've talked for about 20, 25 minutes now, but we haven't even talked about fixed. So with, with the mobile launch of Blackwell, we also have the fixed launch. There's two really big things that I took away from Jensen's keynote around kind of the fixed. I want to see if we have the same takeaway. What are the two big things? And one of them is around VRAM and one of them is around Wanted. So what were kind of. I would love to hear your perspective on those and why that those really matter. Because they do.

Logan Lawler (23:09)

Yeah, yeah, yeah, of course. So we are for the first time offering 96 gigabytes of GPU memory in a single GPU, which is absolutely nuts. And I can't wait to see what people are going to do with this. We're releasing technical demos to kind of showcase some examples. But that is more GPU memory than has ever been available in one of our tests, dedicated desktop workstation GPUs. And it's going to allow for integration of highly advanced 3D and AI pipelines, simultaneously loading massive large language models directly into your system. So the memory boost is something we're really excited about, in addition to the huge performance gains we're driving through. You mentioned differences in wattage. So for the RTX Pro 6000, which is the flagship GPU I'm talking about with the 96 gigabytes of GPU memory that we announced, there's two flavors. There's one that is a 600 watt. So you know, basically it can pull more power to drive more performance and provide max possible performance out of that GPU. And then there's a 300 watt version for more power efficiency and, you know, accommodating different workstation form factors and things like that, that still provides a ton of performance and all that great memory. It's just kind of a difference in whether you want to go extreme extreme on the performance side or get really good performance.

John Dellabona (24:49)

So, yeah, do you want amazing or mind blowing?

Logan Lawler (24:52)

Yeah, basically.

John Dellabona (24:53)

Exactly. So, I mean, and, and I, I think it's crazy. And I, and I'm not super familiar with the, the server side GPU, the H1 hundreds, H2 hundreds, but I think they're in like the 140 range, I believe in terms of like GPU memory per, or approximately like, and I, I think it's a game changer. I can't tell you how many times that in some of the projects that I work on, where, you know, it's like, you know, out of memory and I, and I run the, you know, ADA 6000, I don't do anything crazy. But, you know, I was recently playing around with the Nvidia Cosmos model and generating kind of 3D worlds and stuff like that. And it got to the point where when I kind of took up the steps and the fidelity a little bit, I could generate, but like it would run out, like if I tried to do anything too long, et cetera. So that is an absolute game changer. I mean, I don't think people understand that even if you're running any sort of AI application, is that on and not talking about ADA 6000s, but even other GPUs, whatever you're using is that you might be running Llama, but it's such a quantitized version of it, you're not getting like the full performance, the full knowledge graph, you're not getting any of that. And now you're going to be able to. And one question I have about the 600 watt and I think Dell will have some announcements around that coming out at gtc so I don't want to spoil the party, but this is a more performance question is that you know the GPUs, you know the 6000 and I think you said 6000x equivalent architecture specs minus the wattage. If you were to take them, it have from the test that you've seen in the benchmarks. Have you seen does the 600 watt perform 2x higher because of the wattage you're getting or does it fall somewhere in between like 0x to 2x better because you're doubling the wattage?

Logan Lawler (26:51)

Yeah, you know, we're, we're still in the middle of the benchmarking stuff.

John Dellabona (26:54)

Fair enough, fair enough.

Logan Lawler (26:56)

So I don't, I don't have exact numbers to share but you can expect a big performance jump between the 600 watt and the, and the 300 watt and we will make benchmarks available so everyone can see that. Amazing, amazing performance jump.

John Dellabona (27:10)

Exactly. I mean and it really, it really depends on the workflow. I mean if you are using you know, the ADA now and you don't want to run out of vram, you don't want to do these things. Look at the, the higher wattage thing because it makes sense. I know that some of the announcements that we'll have at gtc, you know, starting in a couple of, you know, Dell Pro Max towers will accommodate this bigger, bigger GPU which is going to be fantastic. You know, whether it in some that are going to hold multiple. So I mean it is really in essence a, a small server, I mean really a small server running at your death side. So we've been on for about 30 minutes. We're going to start winding it down. But I, what I would like for your well to leave with is, you know, after the keynote, you know, obviously bunch of announcements head is still reeling. You know what are kind of I guess the couple of big takeaways that you know from the, the Black Low launch for, for delpromax and workstations in the market. What do you think are the probably the two or three really big things that you took away from the keynote. Like what are the things that's got you, John? Like you're like oh my God, this is amazing. This is going to revolutionize X. Like what are Those two or three big things that you're most excited about?

Logan Lawler (28:26)

Yeah, I think, number one, this, the impact of AI is widespread and it's going to continue to be infused into every industry and continue to be adopted in all these applications. So it's really exciting that we have GPUs at the desktop or at the laptop that people can take to go today, go out there and accelerate their applications and have that kind of dynamic flexibility between the data center and their workstation. You know, they can, they can attack their workloads locally and prototype and then, you know, go out to a cluster and scale out performance there. So I think that's one thing. The other is just how fast AI is developing in itself. You know, we're now at this point where AI is reasoning, requiring additional additional tokens and additional inferencing throughput just to, you know, reanalyze the response before it gives it to you. And then, you know, when you're talking to an LLM and give you very thoughtful, deep answers. Right. So we're seeing inferencing computation increase and we're really excited that, you know, Nvidia GPUs are perfectly positioned to accelerate inferencing workloads across applications. And then I think there's also this thing about physical AI that we're venturing into and is rapidly progressing is this world where we prototype and we develop things in the virtual world. Whether they're new products or robotics or autonomous vehicles. All of these things need to see as if they were a human in the real world. So our RTX technology that fuses photorealistic graphics with AI computer, lets all these things happen in a virtual environment before it's ever deployed physically. And so I think we're in this really exciting time where the future is happening now. And all of these things that we dreamed would be possible in science fiction are starting to come true when it comes to training and deploying robots, or training and deploying autonomous systems, or identifying new, new breakthroughs in health care and things like that. So there's, there's so much going on and I'm, I'm really excited to work at Nvidia and, and be in the center of all of it.

John Dellabona (30:50)

Yeah, I mean, you're right. I mean, I mean, yeah, I mean, it's things that I kind of thought were never possible are coming together because, you know, of advances in AI. But I mean, a lot of that has been, you know, empowered and enabled by, you know, RTX Pro, kind of Blackwell or will be advanced by you know, RTX Pro Blackwell GPUs from, you know, Nvidia and some of your SDKs like Omniverse, which you were referencing, and some others. So, you know, I think you're right. It's an exciting time. So for those that are seeing, I mean, big takeaway, if you have the chance, definitely go back and watch Jensen's keynote. Tons of great stuff, very inspiring, very great stuff. And, you know, look for some announcements that we're going to have our next episode. Not going to spoil the party, but is about a couple of new Dell Promax products, which are very exciting. So with that, John, any. Any final thoughts before we. We sign off and wrap up?

Logan Lawler (31:52)

None for me. This was a great conversation and I'm really happy I was able to, you know, jump on and talk about the technology and happy to do it anytime. So thanks so much for having me.

John Dellabona (32:03)

Yeah, of course. No, we appreciate the time. And, you know, honestly, I'm going to make a joke. Alan picked the right person. This is a perfect conversation for you to wedge in. I love it.

Logan Lawler (32:12)

Thank you so much. I appreciate it.

John Dellabona (32:13)

We're going to sing your praises. This is great. So with that, you know, this kind of wraps up episode two. So, you know, as we get into, you know, kind of future episodes, we're kind of setting the stage right now in terms of kind of what is Dell Pro Max, what has changed on kind of the Dell technology side with workstations. This episode's kind of bringing it together with the Nvidia piece around, you know, pro RTX, Blackwell GPUs. And then as we get into future episodes, obviously we're going to be talking about product and stuff like that, but as we advance, this is all about kind of workflows and how Delpro Max is, you know, reshaping, you know, which is being powered by Nvidia RTX Pro gpu. So, you know, with that, more to come, more workflows, very excited about what's to come. So, you know, if you're a gtc, you'll probably see me around with that. Going to go ahead and sign off and we'll see you on the next episode. What you want?

Logan Lawler (33:27)

This podcast was produced in partnership with Amaze Media Labs.