amx

https://github.com/corsix/amx indicates that Apple's AMX supports up to 64-bit FP, but I don't see any performance metrics. They also have the ANE, which is the low-precision ML-focused accelerator.

Peter Cawley has a good write-up on the undocumented M1/M2/M3 AMX instructions: https://github.com/corsix/amx

There are a lot of undocumented parts of the Apple CPUs, for instance AMX. All such undocumented features can normally be exploited only by the libraries and applications provided by Apple themselves, but not by the applications and libraries written by other parties, which are disadvantaged. This is the same mechanism by which Microsoft has eliminated the competition for Microsoft Office, which used undocumented Windows APIs so that the products of any other vendor could not keep up with it, especially after the launch of any new Windows version. Now one can find some more complete documentation for the Apple CPUs as the result of reverse engineering work done by various people, but after each introduction of a new Apple CPU model the reverse engineering work may need to be done again. Examples: https://github.com/name99-org/AArch64-Expl

Note also that Apple SoCs also have a matrix unit, called AMX, which is designed specifically for matmuls. You could expand the math throughput of the CPU quite dramatically by taking advantage of it. I don’t think there’s any official public documentation on it, but here’s a GitHub repo that documents it independently https://github.com/corsix/amx

Apple has the NPU (also called Apple Neural Engine), which is specific hardware for running inference. Can't be used for LLMs though at the moment, maybe the M4 will be different. They also have a vector processor attached to the performance cluster of the CPU, they call the instruction set for it AMX. I believe that that one can be leveraged for faster LLM inferencing. https://github.com/corsix/amx

I am talking about the matrix/vector coprocessor (AMX). You can find some reverse-engineered documentation here: https://github.com/corsix/amx On M3 a singe matrix block can achieve ~ 1TFLOP on DGEMM, I assume it will be closer to 4TFLOPS for SGEMM. The Max variants have two such blocks. Didn't do precise benchmarking myself, but switching Python/R matrix libraries to use Apple's BLAS result in 5-6x perf improvement on matrix heavy code for me.

I honestly feel cheated by intel with my 13th gen cpu. They took away avx512, even though the P cores are capable of it, but also did not ship AMX with it. I'm honestly more interested in Apple's AMX "co-processor", in my tests it often beats the GPU in ML tasks with smaller batch sizes. It's really cool. I hope it becomes more common for training/inference/numerics capable accelerators to be included in consumer hardware. Apple's AMX is really under-documented, while the instructions were reverse engineered, Virtually no benchmarks are available comparing current chip generations, models and variants. https://github.com/corsix/amx https://github.com/philipturner/amx-benchmarks Curious to see how Intel's AMX stacks up to Apple's AMX in real world perf.

Visible in the unofficial documentation for AMX instructions too - M2 only bf16 functionality - https://github.com/corsix/amx/blob/main/matfp.md This matfp instruction computes an outer product and is a kernel for matrix multiplication.

Confusingly there are 2 mechanisms to do matrix operations on the new apple hardware - AMX (https://github.com/corsix/amx) - and the ANE (apple neural engine) - which is enabled by CoreML. This code does not run on the neural engine but the author has a branch for his whisper.cpp project which uses it here: https://github.com/ggerganov/whisper.cpp/pull/566 - so it may not be long before we see it applied here as well. All of this is to say that it actually could get significantly faster if some of this work was able to be handed to the ANE with CoreML.

really? seems pretty well documented here: https://github.com/corsix/amx

You are correct, in that those are the four My understanding is that the AMX is more tightly wound with the CPU, ultimately being accessible via an instruction set (https://github.com/corsix/amx), and it is useful if you need to do matrix multiplications interleaved with other CPU tasks. A common example would be a VIO loop or something where you want that data in the CPU caches. The GPU and Neural Engine are not that – they take some time to set up and initialize. They also can parallelize tasks to a much higher degree. The GPU is more generalizable, because you can write compute shaders to do anything in parallel, but it uses a lot of resources. I'll have to check out the PR to see how exactly the MPS implementation matches up with the task at hand, because you could also consider writing Metal compute shaders by hand. Even if the perfo

If someone were to build a Linux system with proprietary, private (undocumented) ISA extensions, how would they do it given Linux is open source? Are there any examples of this being done? Would it be possible at all? I got inspiration from this (https://github.com/corsix/amx) and I wondered if someone has done it before on a Linux-based system. I understand a userspace library could be created to access those instructions from userspace, but how would then they be implemented in the kernel? Through a proprietary kernel module built using a custom compiler? Or is that not needed at all and the library could just run on the processor taking advantage of the proprietary extensions?

Which compiler would be required for this (https://github.com/corsix/amx/blob/main/aarch64.h)? I understand the limitation is not at the OS side, as nothing can be done there, but at the compiler-side (I mean that the Apple-supplied compiler doesn't compile against the AMX instruction set, so you'd need a compatible one that, I understand, doesn't exist). Or is it just undocumented and you can actually get it to work with a Standard xcode and macOS installation given the headers provided?

Interesting but the repo homepage itself is probably a more helpful starting point. https://github.com/corsix/amx