11:08
Yeah. You saw the title. You know what this is about. 192 GB of VRAM. It's the Evo X5 from GMK Tech. Gorgon Halo AMD. Let's take a closer look. The release of this is suspicious. Very, very suspicious. You see, AMD has released Gorgon Halo, and so partners like GMK Techch can build systems around the Ryzen AI Max 495 Plus from AMD. This particular configuration has 192 GB of VRAM. It's still available with 128 GB of VRAM and that sort of thing. But it's interesting because AMD is being very bold here. Nvidia's best platform like this has 128 GB of VRAM. This has
192. Yes, naysayers will say, "Well, but the platform's only 273ish gigabytes per second of memory bandwidth. Isn't it fundamentally limited by the amount of memory bandwidth that it has?" Yes, and I completely agree with you. 128 GB of memory seems absurd for 273 GB per second of memory bandwidth because you're taking 500 milliseconds to read all memory. I mean, isn't fundamentally all AI processing gated by the amount of memory bandwidth you have? And yet, and yet, the frontier labs have come up with ways of dealing with relatively limited memory bandwidth. things like Ingram offloading where it might actually be viable to reach 180 billion parameter model running on this thing where it's a
little easier to do in 192 GB of memory versus 128. We need to take a look at that. This is my very first Gorgon Halo machine. Come on. GM K does pack their stuff. Well, a welcome envelope. Hello and welcome. Port layout multiple languages explaining how it works. Okay, cool. I like this form factor a lot. Kindly note for faster setup. Avoid connecting to LAN initially. Ah, comes with Windows. We have two USB 4 V2 connections at the back. HDMI combination headphone microphone. Two type A ports and dual 10 GB LAN. GMK Tech has learned. At the front, we've got two USB 4 ports. This is the 40 gigabit version with 80 in the
back. Two type A ports and a mode button as well as a tiny little reset button, another combination headphone, microphone jack, and a power button. Now, this can be used in a tower or desktop configuration. It's a nice metal case. Got plenty of breathability on the top and bottom. No vents on the sides. Lots of vents at the back. One really super minor thing is that if this is a Kensington lock port, this square cutout on the back is part of the plastic on the back. It really should be part of the metal frame instead. It's a minor
oversight, but it would be a welcome design improvement for those of us that have to endure the uh the unwashed masses of uh university here in the United States. Maybe the high trust society collapse. I don't know. That's a story for another day. In the box, you've also got a bundled power cord and an HDMI cord, which I'm not going to bother unboxing because I got to send this back. And here we have our light on 330 watt power brick. What's up with that? Don't forget, a fair bit of your power budget can go to those USBC ports. Now, as far as Windows goes, AMD has put a lot of work into rocking on Windows and the developer experience on Windows, including RDNA and especially around RDNA. So you want to check out Lemonade
and the Lemonade Server. Lemonade is technically multiplatform. We'll we'll talk about that in a minute. There's a lot of options for Windows and Windows development. Python, especially in Windows, is a first class experience. Is probably a better developer experience than CUDA on Windows natively. There's some asterisk with that. Yeah, I get it. But the Windows development experience is something that AMD is spending a lot of time on. I, however, am super excited for this platform. I'm going to skip straight to Linux. Oh man, one of the things I should do once I get this up and running is uh you know the Starcraft noise is like we require more minerals.
I should take that voice train AI voice on it and then we require more economic development because I don't know. I think only people employed in the industry it's like have you noticed the price of groceries lately? You thought 128 gigs of VRAM was opulent? Try 192 the capital O in opulent. For our bio story, there's a couple of really interesting things that I noticed. First, it has Dash support. So, this platform should work for AMD's answer to Intel VR Pro for outofband management. You can do out of band management with this platform, although you should confirm that with GMK tech, but it looks like the platform supports it. The AMD like PBS CBS menu is not
exposed, but you do have control of uh Sri and resize bar and some other options like that. You can also set the fan performance mode. You can set it to performance or quiet mode. In quiet mode, it is inaudible on the desk. It is dual realtech nyx. As I said before, you can boot off of the network that is supported by the UEFI. There's a lot of interesting UEFI options in the BIOS in general. It's nice to see that it's not locked down in that way. And it's really not a custom BIOS. It's just you get what you get. They didn't paper over anything trying to make it special.
It's just raw access to the BIOS. This is good. Other mini PC manufacturers could learn from that. Well, it booted Windows. It dishonored my request to boot from USB. Probably because secure boots enabled and I've got Vento on there, but hey, I'll allow it. The first spec that I wondered about is that USB 4 V2 controller. Yes, that's the Intel JHL 9580 80 gigabit. That is much easier to work with than the 40 gigabit, which really only gives you 20 gigabit each way. when we're talking about networking and anything else, you can do the RDMA primitives and you can like there's a lot of research down that rabbit hole, but the built-in USB4 controller, nobody has really been able to unlock greater
than 20 Gbits each way on that interface when you're using it for clustering. Meanwhile, 80 Gbits each way is kind of possible with the RDMA primitives, but the software for that is still h it's a little halfbaked. Our configuration here, as I said, is 192 gigabytes. But what I didn't say is that it's 8533. This is the fastest configuration that we've seen. Uh, other than that, it's really not a ton of upgrades over the original Stricks Halo. It's like basically the same Stricks Halo CPU and compute platform, but more memory and faster memory depending on the Stricks Halo platform that you're you were looking at. So, 16 cores, 32 threads, which is really important. We'll come
back to that in just a second. It has three 2280 M.2. This configuration shipped with a two TB M.2, but you got two empty 2280 M.2, so you can add whatever. The dual 10 GB Ethernet, as I said, is realtech. It has a discrete TPM 2.0 chip. So, they didn't go for AMD's FTPM. They added a discrete TPM, and that makes more of a difference on Windows than it does on Linux, but that can have positive performance implications on Linux for complicated reasons I'm not going to get into here. like the discrete TPM in server scenarios generally is a pretty common thing in enterprise deployments and continuous 24/7 use as opposed to the FTPM. Our two TB SSD is a Micron M310, so OEM class M.2, but pretty good performance. When you're doing a lot of
continuous writing, it seems like the SLC right cache is on the order of about 200 gigabytes because the right speed uh does tank pretty hard once you've written 200 GB continuously. The wireless Wi-Fi 7, it's a MediaTek 7925. That does also support Bluetooth 5.4. Now, the power situation, the power envelope is pretty interesting here. 54 watt on the silent configuration, 85 watt for balanced, 120 watt sustained with a peak power draw on the CPU side of up to 160 watts. So when you're running inferencing, a bursty workload like that, the CPU can burst up to 160 watts. And GMK Techch has done a lot of work tuning this platform. Now, when I was doing experiments on this, I had a
little trouble with sleep and wake when it was in this high performance mode. So, I would recommend that you disable uh suspend or going to sleep on this platform. And it was a lot more stable for me after I did that. What does 192 GB of memory get you that 128 GB of memory doesn't get you? The main thing for me is more context and a little bit more concurrency because you can maintain more context across more threads. Doesn't really give you more memory bandwidth. It's the same number of channels. Okay, the memory is clocked a little higher than some Stricks Halo configurations. And because it's clocked a little higher, you get a little bit more bandwidth, which helps, but mainly it's context and running larger models.
The really like the cost of memory, the cost of the extra 64 gigs of memory is so high. It is so hard to justify being able to run uh modern architecture models which have sort of uh the idea of different memory tiers built in. Like if you look at Quinn 3.8 flash next which uses Quinn's next generation architecture. Quinn 4 is not really out at the time that I'm doing this video. It has weights that are designed to run from VRAM and it has weights that are designed to run from slower memory. And that can be storage, that can be system memory. In our testing here, it really didn't make a huge difference in performance, almost none, when you were running the Deep Seek V uh 4.1 flash weights, the
Ingrams as they're called, from the SSD, which freed memory. So, if you're running a 4-bit version of that, a 4-ish bit version of that, you're only going to use 62 GB of memory before you get to your KV cache and everything else. So you've got almost 120 gigabytes left for the rest of the system, any other jobs that you're running, other models that you might be running at the same time and KV cache. So that helps you with concurrency up to, you know, like four or eight workers to just maintain this huge KV cache. It doesn't mean that they can all be busy at once. you don't necessarily get scaling like that. But it is nice that you have the option of having all of those workers with, you
know, the full context that the model supports or maybe even using yarn to extend the context that the model supports and almost all of the testing that we did is using H hallogen because H hallogen runs really fast. The limitation here is 273 GB per second. It's a problem. But with H hallogen and reformatting how the models are stored, multi-token prediction like two or three tokens predicted at once and validated through the engine becomes more possible. I still have a lot of reservations about using um an overly quantized model like the result from that is not as good. Sometimes you can make that up by having a good harness.
Certainly Nvidia in their marketing messaging is saying that like you can have a really good harness, don't worry about the loss of model quality. And I don't I don't 100% buy into that. I would much rather be running, you know, like an 8-bit model if I can as opposed to a 4-bit model or a mixed 4 and 8 bit model. And H hallogen for that mixed 4 to 8bit sweet spot uh with 192 gigs memory, it does open up a lot of possibilities. It doesn't it's not you know the cost benefit because memory is so expensive between 128 and 192 it's a tough cell it's a tough cell for me but from 64 to 128 yeah because you can run some of those Frontier models with a lot of context and another 64 gigs of memory
beyond that it seems like I mean normally I would say this future proof but in the future the cost of memory may come down significantly. Let's take a look at our performance numbers and you can see more concretely what I'm talking about. So, [snorts] the headline is Quinn 3.8 Flash Next on Gorgon Halo is at performance par with the DGX Spark. Actually, it's even better. If you look at the graph, Flash Next AMD Box leads the way at every depth, 1,400 tokens per second. It's twice as fast as DGX Spark. And that's true. That's true at the time that I'm doing this. However, VLM has an unmmerged patch that makes DGX Spark about twice as fast. So, with that, you then get performance par. And
Quinn 3.8 next is not a small model. Uh, it's it's it's quite large. The secret here and the difference is down to software. and software is a big part of this conversation, but I'm going to talk software on the back end where it's not a really a world on the AMD platform where you can just go and grab VLM and have the best possible performance experience. Now, to get this done, we are relying on an engine called Halogen and Halogen's worth a look and it's worth a conversation, but for now, I just want to talk performance. Let's get through the performance because there's more than just Quinn 3.8 next. And the Quinn 3.8 next numbers are quite good on Gorgon Halo and by extension stricks
Halo as well. So if Quin 3.8 flash next is that fast Quinn 3.8 27B dense it's 27 billion parameter model it's much smaller it should also be at performance par right not so much and this is also using H hallogen now there's a H hallogen server for Quinn 3.8 next and there's a H hallogen server for Quinn 3.8b may be and the H hallogen server for Quinn 27B dense is one of the fastest ones out there that you can get. But before we talk multi-token prediction and anything like that, we can talk about like the file formats like the AMD H hallogen format. And so that's this version of Quinn 3.827B. It's the same quant versus the NVFP4 from Nvidia. And if you look at it on disk, it's like, oh, it's
27 20 gigabytes versus 33 gigabytes. I thought that was weird. It's like, okay, the AMD system is going to be at a huge disadvantage. It's like no the H hallogen format here we've got 12 gigabytes of extra stuff here that just helps with uh read at prefill. So this is not what you think. Um the actual load into memory is only about 1.07 times larger than what you get with the NVFP4 format. So this is pretty close to par. If you prefer a bits per weight calculation, I did that at 6.4 bits per weight on the AMD side versus 5.99 on the Spark side. So the disadvantage is very slightly on the AMD side. And for prefill, it looks brutal. And I think this is because there's more room for more optimization on this side.
But also understand this 27b dense model is doing more math per token than Quinn 3.8 flash next. So Quinn 3.8 27B is more math per token almost par. You if you have a simplistic view of it, you might think, well, it's only 27 billion parameter model. Yes, but it's a dense model. And so, you know, 2,300 tokens per second on the DGX Spark platform and Stricks Halo, Gorkin Halo with the software platform can't compete with that. Tried VLM and H hallogen. But look at you look at project Maya. It's like, oh, if we don't have all of the bleeding edge optimizations in VLM, it's like, well, then the performance is very similar. uh or if we don't have all the optimizations, you know, as far as like
Project Maya goes, like it the performance is very similar between the two platforms. And the fall-off on the AMD platform up to 32K of prompt processing really is pretty good. Now, in terms of token generation, this is much closer, much closer in terms of token generation. You know, 14 15 1.5 tokens per second versus about 16 1.5 tokens per second. We're within spinning distance there. And you might notice the green line where it's faster doesn't go all the way across in terms of generated tokens. And it's like having a little bit of problems with the engine and optimization. And that's a problem that can be overcome. But just in terms of like decode performance, not prompt processing, but actual decode
performance on the 27B model. This is parody. This is parody in my opinion. And you can see that as you get pretty deep, you don't really get a ton of performance fall off. You know, 10 tokens per second versus 12. A little bit of variance here is to do with what you're generating. Are we generating pros? Are we generating code? I tried to keep this as like for like between the AMD and Nvidia platforms as possible. And it's very slightly slower, but it's it's basically the same for both of these. And then if we start talking MTP and Dlash 2, that conveys advantages to both platforms. And so you can go from the not usable in my opinion 10.7 tokens per second on 27B dense to 34 tokens per second on Quinn 27B dense which is
pretty phenomenal. Now this Gorgon Halo platform is 192 GB memory though. You want to use a ton of memory. You can run the FP16 version of that, I suppose. And 27B dense, you're going to be using north of 64 GB of VRAM in an FP16 configuration, but in this configuration at 34 tokens per second, it's a reasonable balance of accuracy and smartness of the model. Like it hasn't been quantized into oblivion, quantized into uselessness, but also like 34 seconds uh 34 tokens per second is I think the lower floor of the performance that you want from a model that you're going to use agentically. And the concurrency here is also pretty good. And just for the sake of completeness, if you check out the forum thread, yes,
I did explore VLM on stricks halo with the relevant patches like I didn't just try to do this blindly. And yeah, it was the RDNA hybrid W4A16 linear kernel. Like there's a lot of work that is going to gone into Rockom and Rockom 10 to help with this kind of thing. And for this test, we were using Rockom attention to the Triton paged fallback, but this is on top of the H hallogen testing as well. So, you really have to, we'll talk about this in the software section, but yeah, I tried to keep it as close as possible, but the prompt processing, you know, Spark takes the lead here with this particular extremely popular model. You got 192 GB of VRAM, you want to use it, right?
Well, GLM 5.3 Flash. This is a 321 billion parameter model. And if you look, there are resources on the internet that will tell you can run this on 128 gig DGX Spark, but it's different. It's different and it's a lower fidelity version of what I was able to run on Gorgon Halo at 160 GB 160 GB model versus 121.69 GB model which if you're going to run this model on Spark you do need the latest BIOS update which lets you take the reservation from the GPU from 2 GB to 512 megabytes. Every every bite counts on this whereas you know with 192 GB you've got a comfortable margin. And here I think the AMD platform starts off pretty strong. You know, 232 versus 316 tokens per second when we're talking about prefill. However, as your uh
prefill size grows, you know, 32K uh it doesn't look so good. 140 tokens per second versus 736. However, at the time that I was working on this video, there is another update. There's another update for GLM 5.3 to run on the AMD platform to improve the prefill speed and get it closer to, you know, 350 400 tokens per second. So somewhere in the neighborhood of 50 to 70% of the speed of Spark and remember Quinn 3.8 Flash Next was basically at 100% performance par which is just mind-blowing. It depends on the model, it depends on the software. You can kind of see the state of things as they are now at the time that I'm working on this video. Yeah, definitely room for improvement for GLM
5.3, but GLM 5.3 at this fidelity is unrunnable on GGX Spark. GGX Spark has the advantage here in that this is a lower fidelity model. So, the gap here is not quite as much as the graph would imply because it's not truly an applesto apples comparison. It's like Granny Smith apples versus Red Delicious. I mean, they're both apples, but like they're not exactly the same. There's a it's a lower fidelity on the Spark platform because it had to be quantized to fit 128 gigs of memory. Even despite that disadvantage, the decode performance on these platforms is basically identical. It looks like there's a slight advantage to Spark. But remember, it's also slightly smaller. So I would say these numbers
are basically equal. And effectively the AMD Gorgon Halo platform here is 3.84 bits per weight, which is quite a bit more than the two on DGX Spark. Deepseek is also absurdly popular and it is another different software story on the AMD platform. What I tried to keep constant between the two platforms when we're talking about DeepSeek was the GGUF artifact. So, it's DS4 Dwarf Star and we're streaming, you know, 9 GB of 340 GB and speculation was off on both just so we can get some baseline performance numbers. Our cold start took a long time on Gorgon Halo, but we can ignore that. Like the cold start I don't think is really all that important. uh 651 tokens per second for prompt processing versus 576 tokens per second on DGX Spark. So
the AMD is doing faster prompt processing here and once again similar token generation speeds for DeepSeek on both these boxes. So once you get past the prefill token generation is basically the same. Uh secondarily I was concerned about okay how do I know if it's being as efficient as it possibly can like is the bottleneck really truly uh memory bandwidth and the answer is basically yes uh from 70% memory bandwidth utilization to slightly more than 100% this is like CPU cache or something this is something funky going on inside the H hallogen engine but yeah basically uh Nvidia's had a lot of tuning to make sure that they're using 99 to 100% of their available memory bandwidth when you're doing like forl
like tests like is not all software is like that. In the worst case scenario on the AMD platform is about 69% versus about 66% on the Nvidia platform, but this is just a quick sanity check. More details about that are in the forum thread. So, there's a few more benchmarks there, but yeah, Gorgon Halo's holding its own and the software you know layout here is way different now than it was even 3 months ago and is in unrecognizable from a year ago. And that's part of why I say it's like the hour is later than most people realize on the CUDA mode because H hallogen and strict Halo toolboxes are not official AMD projects. They're just things that randos have done. And how is it possible that randos have done these
things? And it's well, it's because the platform is open and the platform is well documented. and AI has crossed the Rubicon where a determined developer, somebody that knows what they want to accomplish and the only reason they're not accomplishing it is because they didn't have a dozen people working underneath them um is less necessary than it was a year ago. It's like you can direct the AI. It's like here's the specification. It's like I know you're doing the math colonel part of this wrong. Llama.cpp I left llama.cpp out of some of the descriptions, but llama.cpp actually lands at an interesting place with Quinn uh 27B. It can actually beat H hallogen for decode and the
decode performance is a little bit better, but its prompt processing is worse on 27B. And so I think the prompt processing with 27B, you want to sort of get all that you can possibly get. But H hallogen, there's a version of it for Quinn 3.8 27B and then there's another different version of H hallogen from the same guy for Quinn 3.8 Flash Next. If you want to run some other model that's not one of these two things, then that's not exactly H hallogen. If you follow this channel for a long time, you know that I always recommend going to the AMD AI playbooks site, which has some good recipes for getting up and started. And this is true with Gorgon Halo, this GMK Tech Gorgon Halo. You can go here, you can
put things together, and it's amazing. Uh some of these recipes, a lot of these recipes will use VLM and or will use, you know, uh bone stock rockum. And if you do that, you're leaving a lot of performance on the table. But I was delighted when I saw that AMD is also using uh like the stricks Halo toolboxes recipe, like the Deep Seek V4 Flash with DS4. It's like, yeah, let's embrace Dwarf Star. Let's embrace stuff in the community. Let's, you know, let's go to town. And so, you know, like Donado Capella that has, you know, I interviewed at AMD's Advancing AI and not just Donado, but also Carlo Pasqual, who I met also at AMD Advancing AI, working on Rockom. Uh, it's the MXFP4 Rockom extension stuff. Like, these are
some of the best ways to get high performance on the AMD platform. AI toolbox cockpit. You just run that. You've got a guey. you pick and it's like, "Oh, I want to use llama.cpp. I want to use dwarf star. I want to use vlm. I want to use comfy UI. I want to do h hallogen flash." Boom. You can install it and you're up and running. But this is a community project. But this is a community project uh from what I understand that AMD is directly supporting. Whereas Nvidia wants to get their arms around everything. They want to they want to control the experience. It's an exercise to the reader, you know, the pros and cons there. Like the pros on the software side is that you know the best argument that I could make for Nvidia is that AMD had the hardware almost a year
before DGX Spark launched. But the software of the last month or two has made the hardware in October of 2026 infinitely more relevant than it was in October of 2025. Um the other thing with the software and the hardware aspect of this is CPU cores. And this is this goes unstated. this almost controversial maybe these 20 ARM cores that are in DGX Spark, they're good. They're really good. They're shockingly good ARM cores, but the 16 x86 cores that you had in Stricks Halo and now in Gorgon Halo are better. Those 16 cores in general for Agentic workloads for the kind of stuff that I'm doing are just a
bit better than the 20 cores that you get from a DGX Spark configuration. Now, there's there's a couple benchmarks you can cherrypick, but like real world for the stuff that I'm doing, it's like here's some containers or here's here's a an agent that's spinning up Docker containers or that's doing code compile or that's just doing whatever. Those x86 cores, those 16 x86 cores, even though you got four fewer of them, they're a lot more powerful and they make up the difference. And so there's a little bit more insight into that in the forum thread, but it's wild how much the tables have turned on the software side of things because you know Gorgon Halo and Stricks Halo was not purpose designed for the AI industry. It
was designed to be relatively powerful graphically as like an APU kind of a thing. Whereas, you know, DGX Spark for SM120 with like the S cache and some of the other stuff like that is purpose-built to do some of these AI workloads and the linear algebra that is involved with this kind of thing. And so, it's wild to me that AMD has had hardware that is this good. It wasn't really designed with the same intentionality that DGX Spark was. And, you know, here we are with this level of performance. So bottom line looking at the benchmarks is like yeah 192 GB of memory is better than 128 and it's basically the same price. Would you rather have 128 gigs of memory or 192 gigs of memory for the same price? But there are other differences in the
platform like the software stuff. The software stuff is a little easier and more consistent and uh maybe a more wellworn road on the Nvidia side than the AMD side. But the fact that we have at least a couple of models and at least a couple of things that demonstrate concretely the platform is at performance par. I mean we know on paper the memory bandwidth is similar between the two, but like what about all the rest of everything else? Well, at least a couple models they are at par and that is incredibly exciting. Do you need 192 GB of memory to take full advantage of all the AI stuff? H basically no. It's still LPDDR5. 128 gigs of memory for LPDDR5 being able to run and do stuff.
Yeah, having more memory to run bigger models for like research or educational purposes, sure, that's handy. But for me, like trying to have a little helper getting things done, I need performance in like the 30 to 100 tokens per second range with pretty good concurrency ideally. And the one thing that I like that the extra 64 gigs of memory buys me is more concurrency and more KV cache to well somewhat more concurrency depending on the circumstances. It's got a little bit of an asterric there but it also gives me more cash and that is nice to have but more cash is less important in a world where there's good Ingram designs. So I don't know the benchmarks cleared it up but also made it worse. So yeah you're welcome. So,
with that performance data, there's something weird and interesting going on with Gorgon Halo and, you know, the GMK Tech box. There's there's not really any faults that I can find with the GMK Tech box. I hope they uh do a good job keeping their BIOS and their platform updated, fix driver bugs, whatever. I like that they're using an Intel Thunderbolt 5 controller. There's a lot I like about this platform. And the Windows side. The Windows side is good too because of the support from AMD, but also what GMK Techch is doing to support Windows users there. No, I think there's something much more interesting going on here. This platform is launching at a time when Nvidia is also launching RTX
laptops that have 128 well up to 128 GB of memory. And DGX Spark has been here for a long time. But it's funny because I told AMD even before DGX Spark had launched, there was an AMD AI event and I was talking to Anous who, you know, is the guy in charge. I was like, "Hey, um, I think people are going to be really excited for Stricks Halo in an AI context and we need to get the software right and they need to have really high speeded networking and we need to put two of them together." And everybody inside AMD said I was crazy. like nobody is ever going to want to network two of these things together. And that was a headline feature for DGX Spark. It remains a headline feature for
DGX Spark. DGX Spark and this platform. It's like they're limited by the amount of memory bandwidth that they have. But it's really a little misleading to say that there's a limitation there because these have as much memory bandwidth as a Thread Ripper workstation. And they are they're tiny. It's just that, you know, AI demands a crazy amount of memory bandwidth and AI demands a crazy amount of memory capacity. And so this bump here is, you know, it's it's 50% more memory, which is great for AI tasks. It comes at a time when memory is unaffordably expensive. And, you know, that's distracting, I think, from the important thing that's actually going on here, which is H hallogen shows what the
hardware is capable of. And H hallogen is limited in that it's really designed to run Quinn 3.8 flash next as fast as it possibly can and it's really not a generic software solution. It is a software solution for this one specific piece of hardware. But for that model with this piece of software using AI assistance to put that together and run that as fast as you possibly can on the AMD APU in here. This uh this signals a couple of things. One, the CUDA hour is later than Nvidia thinks. See, CUDA is a little bit of a moat and everybody looks at CUDA and says, "Oh, Nvidia has the software stack. They have the software ecosystem." And you know, CUDA is the cornerstone of that. CUDA is not and has
not been the cornerstone of that for some time. But the folks that still think that some aspect of CUDA is a little bit of a moat, randos on the internet, not AMD proper, were able to cobble together a software math kernel to take full advantage of the hardware available in the Halo platform and deliver something that is highly competitive with DGX Spark. And in terms of memory capacity and this one specific model exceeds DGX Spark on a couple of key metrics. Yeah. That's kind of interesting, isn't it? 192 GB of memory when everything else is 128 GB of memory. And that's the headline, you know, that Nvidia and AMD want to talk about. And of course, Nvidia wants to focus on the software stack.
And AMD wants to keep your focus on the memory capacity and the story the software story is a lot more interesting if you break down like the H hallogen part of the software story. Now there are other uh folks that are doing really interesting approaches. Loose box for example they've got their system that's based on the same kind of a platform but they also couple it with an R9700 dedicated GPU DGPU. So lots of LPDDR5 plus this thing connected on the PCIe bus. And again, it's still not like what Nvidia is doing with the coherent memory and everything else, but it's interesting that solution cobbled together by third party companies is kind of viable for more narrow use cases
in a lot of ways. The other thing that you look at is like the Agentic workloads. This is another thing that Nvidia does not want you to notice. The 16 cores in here are generally much better than the 20 cores that are in DJX Spark. these 16 x86 cores are beating the crap out of those 20 ARM cores, you know, 10 and 10 ARM cores. Is that going to be true? You know, that's that generation. Is that going to be true in the next generation spark with Vera Rubin? because we know that Nvidia has put a lot more work into the newer ARM cores and I think that you know we have you know Zen 6 is also right around the corner and that is uh that is going to be a very interesting battle and I don't know how that's going to break down but it's
interesting that you know laptop space on this versus an RTX spark laptop you know 20 cores ARM got the ARM goodness Windows on ARM now And this is just quietly 16 cores of x86. And it's not as many cores, but those cores are more capable. Again, sort of interesting when you look at that. It's like, well, wait, x86 is winning over ARM and in both battery life and performance. Huh, that's interesting. At least in that one narrow scenario. Now, ARM can still win in battery life when you're talking about like Snapdragon, that kind of thing, but like that is not there's no AI there. That's not that's a different thing really. So, I don't know. I'm level one. It's fun thoughts. This a fun
early look at the Knuck box from GMK Tech and sort of getting to ramble a little bit. 192 gigs of LPDDR5. I wish this wasn't gosh darn expensive. Existential dread setting in. Well, I got to pack this up and send it back. But you can get a coupon for like $100 or 5% off on the Amazon link. Check out the forum thread. There's a coupon. You can save some money. You want to save money, right? Cuz oh my goodness, the cost of everything. Oh my goodness. Come join me in the forum. Hang out. Let's uh let's keep it together. Together on the forum. I'm signing out and I'll see you there.
