Hacker News Reader

Any improvement like that is from the GPU being given different kinds of circuits entirely. The raw performance difference is not very impressive. In late 2012 you'd get about 8 gigaflops per dollar. Late 2017 gets you almost double that. They're being used somewhat more effectively, but a normal non-neural-net benchmark shows a total performance improvement of... 3x.

They're doing better than CPUs but they're still falling behind Moore's law.

View on HN

> a normal non-neural-net benchmark shows a total performance improvement of... 3x

because they aren't intended to run non-neural-net workloads. you don't judge a dishwasher by it's ability to wash your clothes, do you?

When you're judging whether Moore's law is dead, you need to separate out the improvements from changing the FPU layout and the improvements from lithography. And since the consumer chips don't dedicate a lot of space to neural net units, they're the best baseline.

It's cool that they made some neural net things go hugely faster. But they could make the same fundamental chip on 6-year-old foundries, only slower by a factor of 2 or 3. That really impressive number has nothing to do with Moore's law.

Reply on news.ycombinator.com