On the SDK front, you tried just having an agent reimplement the model you're interested in and just use the weights? I've taken to treating off the shelf implementations as reference implementations anyway, because I can often squeeze out significantly better performance for my configuration and use case by having Codex hammer at it for a few hours.
It's been pretty ad hoc, but my prompts are nothing special. Things I generally do:
1. Top end model on high/xhigh thinking (last time I did it it was Sol xhigh I think)
2. Make sure it creates some representative fixtures of different sizes and sets up a good testing, profiling and benchmarking loop that doesn't require my input.
3. Make sure it has access to reference implementation code
I would too, but sadly that's Khronos' job to organize, and they've had trouble getting American vendors to work together.
It's likely that CUDA will continue dominating until they put aside their differences. The current MLX/MPS/ROCm ecosystems are too fractured to threaten Nvidia.
Yeah talking about the (vendor-preferred) compute part here
Vulkan's SPIR-V dialect is substantially different from the OpenCL one, notably with the former having structured control flow. They're incompatible between each other.
I'm talking about holistic efforts like OpenCL, and standards that would be equivalent to Nvidia's "Compute Capability" versioning.
The basic underlying tech can be agreed on, but Apple/AMD/Intel all have different GPU priorities that limit their ability to agree on a CUDA-adjacent hardware platform.
What do you mean by holistic? SYCL is an open versioned standard that allows for vendor specific extensions. The problem is not that there isn’t a proper standard, the problem is that many hardware vendors - or software developers simply don’t want to adopt it.
Intel (via Codeplay) was handing it out on a silver platter - Nvidia on SYCL, full top chain, and people still wouldn’t want it.
Isn't OneAPI a good example of the problem, alongside Mojo/ONNX/TensorRT? The industry doesn't need a fifteenth competing standard. They need hardware buy-in.
By holistic, I mean hardware architecture cooperation. Nvidia can hold onto their lead forever if GPU designers fight over what a GPGPU hardware baseline looks like. The current ecosystem fragmentation is not competitive, and future fragmentation probably wouldn't work either. I think the fastest way to kill Nvidia would be a hardware consortium.
The problem with OneAPI is naming. It leads people to believe that is another competing standard where in fact is is simply just an implementation of a standard compliant SYCL compiler. If it just had been named “Intel SYCL compiler”, similar to the existing and accepted Intel OpenCL compiler, it would have been easier.
What would you expect the hardware consortium to coordinate on? Unified ISA?
They were maintained by Codeplay - paid for my Intel. Nvidia can make contributions anytime they want, and here is the problem: Nvidia does not want to. Until each vendor starts pitching in with contributing their backend to an open standard, you will have to rely on others doing it for them.
AI will take down Nvidia’s moat. When it becomes trivial to translate CUDA/PTX to HIP, SYCL or Metal, CUDA is no longer the moat, it becomes the intermediate representation.
yeah yeah, "when" an often keyword with AI it seems. As Mr. E. Nigma put it - what always comes but never arrives? Meanwhile the moat deepens and it's build on inertia and laziness and Nvidia knows this really REALLY well.
i swear people who are outsiders here have only clickbait takes; if you've never had to ship GPU code professionally you should just not comment on these things.
the source language has never been the moat. Nvidia sells to hyperscalers. Hyperscalers have armies of kernel authors who have no issue translating shaders by hand (or now with claude). Nvidia's moat is (and will remain for the foreseeable future) the entire stack. you cannot fathom the pain and misery of working on literally any other stack. if you've never debugged a GPU synchronization error or kernel panic due to some GPU firmware bug or fought absolute shit profilers hunting for perf you really have no idea what you're talking about.
this isn't a "pissing contest"? you made a speculative claim in a public forum and i'm challenging your authority to make such a claim. a "pissing contest" would be if i had said i've shipped hundreds of thousands of lines of shader code into prod and thus you clearly have no idea what you're talking about because you haven't (which is also true).
I could post the GitHub URLs of all the shader code I wrote that’s running on countless GPUs right now, but what would it change? I’m still just a random guy on the internet with an opinion that happens to be different from your opinion.
You can simply disagree with me, regardless of my experience (or lack thereof).
RDNA1 isn't good for a whole lot, even flagship RDNA2 cards are a stretch for many things. The lack of WMMA/matrix multiply/BF16 is too severe of a penalty.
The FP16 throughput on RDNA1 is both shader reliant and requires everything to be packed first. Even with 2 or 4 or 1000 cards, you would be consuming all of the available memory and memory bandwidth just packing and unpacking values, and if you really want to dump a hundred billion tokens into making it work anyways, you're only going to find out that even if you bother to sit there ferrying packed values to ram or disk before then issuing the instructions, paying that already severe penalty again when the values then have to be unpacked is so steep of a cost that the 256 BF16 flops/cu/clock's effective throughput is outright lower than simply doing it on a Zen 2 processor. You also don't have INT8 (or really INT4) on RDNA1 so the other RNS/CRT tricks aren't viable.
Sadly RDNA1's VCN2 also lacks actually good x264 bframe encoding support, or even P010 for 10 bit color, so what I'm saying is you should sell them. Used Radeon VII's are like $260, you'll go a lot further with those especially if you throw in a 7900XTX, and then augment that further with a 9070 CRE (you only want it for its int8 cores), and of course 128GB of ram.
E: And sure, that's 3, or ideally 4 GPUs, and a good bit of extra work. But that gets you up to more than halfway to the naive performance of a $15,000 MI300x in a surprising amount of cases, with additional strengths that it lacks. For far less than half of the cost
So far, it has only been tested on my RX 9060 XT (gfx1200), where I’ve used it with CUDA-enabled LibTorch workloads, including long ai training and use.
I also added a GPU scanner / auto-detection system that detects:
AMD GPU model
gfxXXXX architecture
ROCm/HIP installation
driver info
whether the GPU has already been validated by the project
The goal now is to test it on more hardware, especially RX 6000 / 7000 / 9000 cards.
If you have an AMD GPU on Windows and want to try it, I’d really appreciate compatibility reports working or broken. There’s a dedicated GPU compatibility issue template in the repo.
If this is useful to you, a star would also help the project get more testers.
It's unbearable that most LLM inference happens on closed H/W, closed drivers, and closed SDKs.
I would really appreciate your input!
1. Top end model on high/xhigh thinking (last time I did it it was Sol xhigh I think)
2. Make sure it creates some representative fixtures of different sizes and sets up a good testing, profiling and benchmarking loop that doesn't require my input.
3. Make sure it has access to reference implementation code
It's likely that CUDA will continue dominating until they put aside their differences. The current MLX/MPS/ROCm ecosystems are too fractured to threaten Nvidia.
Intel uses SPIRV iirc
They're migrating away from SPIR-V to their own, Intel PISA: https://discourse.llvm.org/t/rfc-upstreaming-the-pisa-backen...
Vulkan's SPIR-V dialect is substantially different from the OpenCL one, notably with the former having structured control flow. They're incompatible between each other.
The basic underlying tech can be agreed on, but Apple/AMD/Intel all have different GPU priorities that limit their ability to agree on a CUDA-adjacent hardware platform.
Intel (via Codeplay) was handing it out on a silver platter - Nvidia on SYCL, full top chain, and people still wouldn’t want it.
By holistic, I mean hardware architecture cooperation. Nvidia can hold onto their lead forever if GPU designers fight over what a GPGPU hardware baseline looks like. The current ecosystem fragmentation is not competitive, and future fragmentation probably wouldn't work either. I think the fastest way to kill Nvidia would be a hardware consortium.
What would you expect the hardware consortium to coordinate on? Unified ISA?
Yes they have implementations on top of CUDA but they're maintained by... Intel. They didn't get buy-in for cross-vendor collaboration
ZLUDA is already doing that, no?
the source language has never been the moat. Nvidia sells to hyperscalers. Hyperscalers have armies of kernel authors who have no issue translating shaders by hand (or now with claude). Nvidia's moat is (and will remain for the foreseeable future) the entire stack. you cannot fathom the pain and misery of working on literally any other stack. if you've never debugged a GPU synchronization error or kernel panic due to some GPU firmware bug or fought absolute shit profilers hunting for perf you really have no idea what you're talking about.
You can simply disagree with me, regardless of my experience (or lack thereof).
The FP16 throughput on RDNA1 is both shader reliant and requires everything to be packed first. Even with 2 or 4 or 1000 cards, you would be consuming all of the available memory and memory bandwidth just packing and unpacking values, and if you really want to dump a hundred billion tokens into making it work anyways, you're only going to find out that even if you bother to sit there ferrying packed values to ram or disk before then issuing the instructions, paying that already severe penalty again when the values then have to be unpacked is so steep of a cost that the 256 BF16 flops/cu/clock's effective throughput is outright lower than simply doing it on a Zen 2 processor. You also don't have INT8 (or really INT4) on RDNA1 so the other RNS/CRT tricks aren't viable.
Sadly RDNA1's VCN2 also lacks actually good x264 bframe encoding support, or even P010 for 10 bit color, so what I'm saying is you should sell them. Used Radeon VII's are like $260, you'll go a lot further with those especially if you throw in a 7900XTX, and then augment that further with a 9070 CRE (you only want it for its int8 cores), and of course 128GB of ram.
E: And sure, that's 3, or ideally 4 GPUs, and a good bit of extra work. But that gets you up to more than halfway to the naive performance of a $15,000 MI300x in a surprising amount of cases, with additional strengths that it lacks. For far less than half of the cost
I’ve been working on a Windows setup that lets CUDA-targeted applications run on AMD GPUs using ZLUDA + ROCm/HIP.
Repo: https://github.com/Speedstu/CUDA-for-AMD-Windows
So far, it has only been tested on my RX 9060 XT (gfx1200), where I’ve used it with CUDA-enabled LibTorch workloads, including long ai training and use.
I also added a GPU scanner / auto-detection system that detects:
AMD GPU model
gfxXXXX architecture
ROCm/HIP installation
driver info
whether the GPU has already been validated by the project
Example:
RX 9060 XT → gfx1200 → RDNA4 → HIP detected → validated
The goal now is to test it on more hardware, especially RX 6000 / 7000 / 9000 cards.
If you have an AMD GPU on Windows and want to try it, I’d really appreciate compatibility reports working or broken. There’s a dedicated GPU compatibility issue template in the repo.
If this is useful to you, a star would also help the project get more testers.