HC2023-K2: Hardware for Deep Learning

Neuromorphic computation has been hyped up for ~ 20 year by now. So far it has dramatically underperformed, at least vis-a-vis the hype. The article does not distinguish between training and inference. Google Edge TPUs link each one is capable of performing 4 trillion operations per second (4 TOPS), using 2 watts of power—that's 2 TOPS per watt. So inference is already cheaper than the 20 watts the paper attributes to the brain. To be sure, LLM training is expensive, but so is raising a child for 20 years. Unlike the child, LLMs can share weights, and amortise the energy cost of training. Another core problem with neuromorphic computation is that we currently have no meaningful idea how the brain produces intelligence, so it seems to be a bit premature to claim we can copy this mechanism. Here is what the Nvidia Chief Scientist B. Dally (and one of the main developers of modern GPU architectures) says about the subject: "I keep getting those calls from those people who claim they are doing neuromorphic computing and they claim there is something magical about it because it's the way that the brain works ... but it's truly more like building an airplane by putting feathers on it and flapping with the wings!" From "Hardware for Deep Learning" HotChips 2023 keynote. ▶ this video This is at 21:28. The whole talk is brilliant and worth watching.

I think your biggest mistake is ignoring the amount of software that runs on top of the hardware. There were times in the past where others had a higher number of FLOPS, but it didn’t make a dent in the usage of CUDA. My standard assumption about anything in tech (and beyond) is that you don’t win on first principles alone (“parallel design is well understood”). It’s the details that make you stand out. Bill Daly gave a great presentation at Hot Chips a few weeks ago that explains how Nvidia’s AI performance was only in small parts due to process improvement. The rest was doing things smarter: matrix sparsity, different number representations and so forth. ▶ this video

Says the hotchips keynote - ▶ this video - data that is moved is about 100x higher power than data that is local You have a good point about intermediate data. This would not apply to all algorithms.