Re:Ex-Googlers left secretive AI unit to form Groq with Pali...
by highd on Friday April 21, 2017 @01:25AM
(#14163283)
I don't think this will have anything to do with training, since it's the TPU team and trying to beat nVidia at floating-point matrix multiplication doesn't seem like a great idea. Run-time can generally be done with integer operations, which is near dirt-cheap these days - I'd think the only interesting things would be operating at server scales or so low-power that you're getting 30FPS on your classification/segmentation what-have-you on your smartphone without impacting battery life. My guess is the former, though I'm not sure what the market size is for companies that need to evaluate their machine learning systems at speeds fast enough to merit specialized hardware.
Deep learning is changing the field of AI and I expect that with a few more breakthroughs in training algorithms and hardware everything will start to change, hopefully for the better. I wonder when we will see ultra low power neural chips in cellphones.
In the 1980s, my team at SAIC built special purpose neural hardware: Harvard memory architecture designed to run backprop quickly. We got 5 megaflops out of our boards and we were happy enough.