Tomeu has successful achieved the target of 3x faster than CPU by accelerating the convolution component of MobileNetV1. He will start cleaning uo the code and upstreaming the Teflon Delegate to the Mesa project.
This is the first mesa implementation of an NPU and we are very excited for the potential applications on our Cottonwood platforms. Together with other performant subsystems, Alta will deliver advanced edge capabilities never seen on the sub-$100 SBC market.
In the next few months, the scope of this work will expand beyond Tensorflow Lite and has the potential to re-invent and accelerate AI and neuro computation methods as developers optimize computation for hardware design. The open source model enables novel approaches for finding the best tradeoffs between speed, accuracy, and complexity.
To run it on your board, see our instructional guide.