News for hackers. Stuff that matters.
Re:Tesla’s Neural Processor In The FSD Chip by _visgean on Thursday September 26, 2019 @06:31PM (#21086296)
To my understanding when you run tensorflow it builds the computational graph in representation that can be processed either on CPU or GPU, so for example it seems that here you would need "compiler" that takes the graph and runs it on the device.. similar to Edge TPU. So essentially you would install drivers and then communicate through standard interface..

But I might be wrong...

Re:Tesla’s Neural Processor In The FSD Chip by pjscott on Thursday September 26, 2019 @06:35PM (#21086330)
Yes, they have a bunch of special model compilation stuff that needs to happen in order for the ML model to run on their hardware, such as weight quantization, and then they'd also need to generate the a program for the NPU so it knows what to do -- what to convolve with what, and so on.