One of my pet peeves around running local LLMs and inferencing is the sheer mountain of shit^W^W^W complexity of compute stacks needed to run any of this stuff in an mostly optimal way on a piece of hardware.
CUDA, ROCm, and Intel oneAPI all to my mind scream over-engineering on a massive scale at least for a single task like inferencing. The combination of closed source, over the wall open source, and open source that is insurmountable for anyone to support or fix outside the vendor, screams that there has to be a simpler way. Combine that with the pytorch ecosystem and insanity of deploying python and I get a bit unstuck.
What can be done about it?
llama.cpp to me seems like the best answer to the problem at present, (a rust version would be a personal preference, but can't have everything). I like how ramalama wraps llama.cpp to provide a sane container interface, but I'd like to eventually get to the point where container complexity for a GPU compute stack isn't really needed except for exceptional cases.
On the compute stack side, Vulkan exposes most features of GPU hardware in a possibly suboptimal way, but with extensions all can be forgiven. Jeff Bolz from NVIDIA's talk at Vulkanised 2025 started to give me hope that maybe the dream was possible.
The main issue I have is Jeff is writing driver code for the NVIDIA proprietary vulkan driver which reduces complexity but doesn't solve my open source problem.
Enter NVK, the open source driver for NVIDIA GPUs. Karol Herbst and myself are taking a look at closing the feature gap with the proprietary one. For mesa 25.2 the initial support for VK_KHR_cooperative_matrix was landed, along with some optimisations, but there is a bunch of work to get VK_NV_cooperative_matrix2 and a truckload of compiler optimisations to catch up with NVIDIA.
But since mesa 25.2 was coming soon I wanted to try and get some baseline figures out.
I benchmarked on two systems (because my AMD 7900XT wouldn't fit in the case). Both Ryzen CPUs. The first I used system I put in an RTX5080 then a RTX6000 Ada and then the Intel A770. The second I used for the RX7900XT. The Intel SYCL stack failed to launch unfortunately inside ramalama and I hacked llama.cpp to use the A770 MMA accelerators.
ramalama bench hf://unsloth/Qwen3-8B-GGUF:UD-Q4_K_XL
I picked this model at random, and I've no idea if it was a good idea.
Some analysis:
The token generation workload is a lot less matmul heavy than prompt processing, it also does a lot more synchronising. Jeff has stated CUDA wins here mostly due to CUDA graphs and most of the work needed is operation fusion on the llama.cpp side. Prompt processing is a lot more matmul heavy, extensions like NV_coopmat2 will help with that (NVIDIA vulkan already uses it in the above), but there may be further work to help close the CUDA gap. On AMD radv (open source) Vulkan is already better at TG than ROCm, but behind in prompt processing. Again coopmat2 like extensions should help close the gap there.
NVK is starting from a fair way behind, we just pushed support for the most basic coopmat extension and we know there is a long way to go, but I think most of it is achievable as we move forward and I hope to update with new scores on a semi regular basis. We also know we can definitely close the gap on the NVIDIA proprietary Vulkan driver if we apply enough elbow grease and register allocation :-)
I think it might also be worth putting some effort into radv coopmat2 support, I think if radv could overtake ROCm for both of these it would remove a large piece of complexity from the basic users stack.
As for Intel I've no real idea, I hope to get their SYCL implementation up and running, and maybe I should try and get my hands on a B580 card as a better baseline. When I had SYCL running once before I kinda remember it being 2-4x the vulkan driver, but there's been development on both sides.
(The graphs were generated by Gemini.)
