I think you're massively under-estimating the difference in specialization. Tesla know the resolution of the data going into the Neural Nets, they know the dept

I think you're massively under-estimating the difference in specialization. Tesla know the resolution of the data going into the Neural Nets, they know the depth of the data, they know the types of non-convolution stages (pooling, ReLU Sigmoid etc). They've probably got a fixed number of layers too. So they're basically completely getting rid of the control logic. That's an enormous chunk of the problem, and lets' remember these are peak performance numbers - the Tesla programming model may be a disaster. Hell, they could probably have bought the IP for the actual computation core.