LAUNCHXL-F28P55X: NPU

WMWMS

Tool/software:

I want to use the built-in NPU of the F28P55X to perform simple mathematical operations (e.g., matrix multiplication). Is it possible to directly invoke the NPU via library functions for accelerated computation instead of model training? If possible, how can I invoke it?

11 months ago

0 Amir Hussain 10 months ago

TI__Expert 5105 points

I looking for the right person to answer your question. We will respond soon.

Thank you

Amir Hussain

0 Yuan Zhao 10 months ago

TI__Expert 4055 points

No. We do not support direct NPU programming at this moment. -Yuan

0 WMWMS 10 months ago in reply to Yuan Zhao

Prodigy 50 points

Thank you for your response! In fact, during the process of motor control, we often use a large number of for loops when writing programs. The CPU usually takes a considerable amount of time to run these for loops. So, I would like to ask if we can utilize the powerful computing capabilities of the NPU to significantly reduce the time spent on executing these for loops?

0 Yuan Zhao 10 months ago in reply to WMWMS

TI__Expert 4055 points

I see. Unfortunately, NPU is not a general purpose computing core. It is more fit for running fixed point neural network layers with multiply and accumulate. If you want to run a neural network inference workload, the NNC compiler user's guide is here: https://software-dl.ti.com/mctools/nnc/mcu/users_guide/ Our top-level user interface for customers is EdgeAI Studio MCU: https://www.ti.com/tool/EDGE-AI-STUDIO