What kind of computations does Ethos-U55 NPU perform? I mean: matrix computation? Trigonometric computation? Multiply-and-accumulate? What types of data is it capable of using? Fixed-point or floating-point? How many bits?
Hi mastupristiEthos-U55 has a dedicated MAC unit which does the computation. MAC unit performs multiply-accumulate operations that are required for convolution, depthwise pooling, vector products, and the max operation required for max pooling., the main reason that you get the speed up in Ethos U55 is that U55 have a fixed function units that end up doing the operators needed for convolutions etc. which makes it really fast. U55 also have SRAM and other bits that help with this and a number of these things are tied in pretty closely with the compiler. The compiler used for Ethos-U55 is not like a traditional compiler but something that knows how to optimize for that core. regarding datatype - Ethos-u support 8-bit weight or 8-bit weight and 16-bit activation. For FP32 models you can do the post training quantisations.
You can refer to the Technical Reference Manual for any further information - https://developer.arm.com/documentation/102420/0200/Functional-descriptionHope this helps.
thank you for the answer.My real question I think has already been answered negatively unfortunately.
In fact I don't really need ML or NN. I do audio signal processing. And so I explore the possibility of using Ethos-U to do general purpose processing (somewhat like GPUs do with OpenCL).
In fact, for numerical accuracy we cannot go below FP32. So I think having 8 bits for coefficients and 16 bits for samples is too little for my MAC needs
You may wish to explore Cortex-M7 and its FPU:https://developer.arm.com/documentation/dui0646/c/Cortex-M7-Peripherals/Floating-Point-Unit
If you are choosing an MCU be sure to verify that FPU is supported, as it is optional, and so not all silicon may provide it.