Next Raspberry Pi CPU Will Have Machine Learning Built In

If it is an instruction thing and not more of a separate coprocessor, it is maybe SIMD on smaller integers for inference (like how AVX512 enables vectorized 8bit integer math, for 64 operations per instruction).

From the article,

"It may come in the form of lightweight accelerators possibly 4-8 multiply-accumulates (MACs) per clock cycle."

My guess is that they'll likely incorporate something like the Arm Neon architecture extension that significantly improves performance in linear algebra and signal processing.

> really just mean "ALUs with additional multiply/add units"?

Well, that's true in general :-)

When CPU companies these days say "machine learning" instructions, they *REALLY* mean 8-bit or 16-bit matrix-multiplication acceleration of some kind. Its very vague, but that's what the machine-learning guys care about.

* NVidia implements Half-float FP16 4x4 matrix multiplication instructions.

* AMD implements INT8, BFloat16, FP16 dot-product instructions.

* Intel implements AVX512 INT8 instructions. I don't know the details however

What will the Rasp. Pi implement? 4 to 8 MAC instructions/clock. MAC will certainly accelerate matrix-multiplications, but not as much as a proper 4x4 matrix multiplication instruction.

Probably not, that would make it very obvious that the vast majority of device sales are not going to the official target audience: education.