Back in June, we mentioned Pete Warden’s port of the Deep Belief image-recognition SDK to the Pi, which used the VideoCore IV QPUs to provide an accelerated GEMM matrix-multiply function. Since then, Pete’s been optimizing his code, and has reduced the time required to process an image to 3 seconds (versus 20 seconds for the baseline ARM implementation and 6 seconds for his original QPU version). Classifying dogs and their balls In the spirit of “leaving a trail of breadcrumbs through the forest”, Pete has written up an excellent summary of his experiences here

Continue reading here:
More QPU magic from Pete Warden