You always could. It's just slower and less efficient because most of what a model query needs to do is matrix multiplication and the GPU is optimized for that.
You're reinventing the CISC wheel that Apple, Google and Nvidia all abandoned.
x86 already has AVX accelerators that are a huge help for LLM inference speed and efficiency. It's just not enough, only the absolute weakest iGPUs are slower than AVX inference. The majority of consumer hardware would benefit from pipelining AI workloads through the GPU, especially on Apple Silicon and other ARM hardware that has borderline nonexistent SIMD support.
ARM will be ideologically opposed to expanding CPU-side SIMD, and x86 is already buckling under the weight of modern AVX (see: AVX-512 issues).
Not sure what this comment is based on. Arm introduced a scalable SIMD whose whole point is to be expandable. AVX-512 works very well, for example on Zen 4 and 5.
8 comments
[ 6.9 ms ] story [ 29.3 ms ] threadx86 already has AVX accelerators that are a huge help for LLM inference speed and efficiency. It's just not enough, only the absolute weakest iGPUs are slower than AVX inference. The majority of consumer hardware would benefit from pipelining AI workloads through the GPU, especially on Apple Silicon and other ARM hardware that has borderline nonexistent SIMD support.
ARM will be ideologically opposed to expanding CPU-side SIMD, and x86 is already buckling under the weight of modern AVX (see: AVX-512 issues).