From reverse engineered information (in the context of Asahi Linux, which can have raw hardware access to the ANE) it seems that the M1/M2 Apple Neural Engine provides exclusively for statically scheduled MADD's of INT8 or FP16 values.[0] This wastes a lot of memory bandwidth on padding in the context of newer local models which generally are more heavily quantized.

(That is, when in-memory model values must be padded to FP16/INT8 this slashes your effective use of memory bandwidth, which is what determines token generation speed. GPU compute doesn't have that issue; one can simply de-quantize/pad the input in fast local registers to feed the matrix compute units, so memory bandwidth is used efficiently.)

The NPU/ANE is still potentially useful for lowering power use in the context of prompt pre-processing, which is limited by raw compute as opposed to the memory bandwidth bound of token generation. (Lower power usage in this context will save on battery and may help performance by avoiding power/thermal throttling, especially on passively-cooled laptops. So this is definitely worth going for.)

[0] Some historical information about bare-metal use of the ANE is available from the Whisper.cpp pull req: https://github.com/ggml-org/whisper.cpp/pull/1021 Even older information at: https://github.com/eiln/ane/tree/33a61249d773f8f50c02ab0b9fe... .

More extensive information at https://github.com/tinygrad/tinygrad/tree/master/extra/accel... (from the Tinygrad folks) seems to basically confirm the above.

(The jury is still out for M3/M4 which currently have no Asahi support - thus, no current prospects for driving the ANE bare-metal. Note however that the M3/Pro/Max ANE reported performance numbers are quite close to the M2 version, so there may not be a real improvement there either. M3 Ultra and especially the M4 series may be a different story.)