XNNPACK VS gemm-benchmark

Why are you writing your own inference code in C++ or Rust instead of using some kind of established framework like XNNPACK?

XNNPACK - https://github.com/google/XNNPACK

Ok, concrete example time! At work, we're going to be using some software which includes XNNPACK, which is a library of highly-optimised operations for doing neural-network inference. This is the sort of thing where people have gone in and specifically tuned for performance, and nope, there's no attempt at all made to have code which is different for Intel/AMD or Apple/Other ARM. What they target is elements of the ISA, like NEON (i.e. ARM SIMD) and SSE, AVX etc. on x86(-64). And Wasm SIMD for Wasm.

It's quite simple. https://github.com/google/XNNPACK for example.

And PyTorch on the M1 (without Metal) uses the fast AMX matrix multiplication units (through the Accelerate Framework). The matrix multiplication on the M1 is on par with ~10 threads/cores of Ryzen 5900X.

[1] https://github.com/danieldk/gemm-benchmark#example-results

Apple Silicon Macs have special matrix multiplication units (AMX) that can do matrix multiplication fast and with low energy requirements [1]. These AMX units can often beat matrix multiplication on AMD/Intel CPUs (especially those without a very large number of cores). Since a lot of linear algebra code uses matrix multiplication and using the AMX units is only a matter of linking against Accelerate (for its BLAS interface), a lot of software that uses BLAS is faster o Apple Silicon Macs.

That said, the GPUs in your M1 Mac are faster than the AMX units and any reasonably modern NVIDIA GPU will wipe the floor with the AMX units or Apple Silicon GPUs in raw compute. However, a lot of software does not use CUDA by default and for small problem sets AMX units or CPUs with just AVX can be faster because they don't incur the cost of data transfers from main memory to GPU memory and vice versa.

[1] Benchmarks:

https://github.com/danieldk/gemm-benchmark#example-results

https://explosion.ai/blog/metal-performance-shaders (scroll down a bit for AMX and MPS numbers)

Yes, there is one per core cluster. The title is a bit misleading, because it suggests that going to two or three cores will scale linearly, though it won't be much faster. See here for sgemm benchmarks for everything from the M1 to M1 Ultra and 1 to 16 threads:

https://github.com/danieldk/gemm-benchmark#1-to-16-threads

There's always been a tradeoff in writing code between developer experience and taking full advantage of what the hardware is capable of. That "waste" in execution efficiency is often worth it for the sake of representing helpful abstractions and generally helping developer productivity.

The GFLOP/s is 1/28th of what you'd get when using the native Accelerate framework on M1 Macs [1]. I am all in for powerful abstractions, but not using native APIs for this (even if it's just the browser calling Accelerate in some way) is just a huge waste of everyone's CPU cycles and electricity.

[1] https://github.com/danieldk/gemm-benchmark#1-to-16-threads