News for hackers. Stuff that matters.
Re:Leela Zero – Go engine with no human-provided knowledge by raverbashing on Thursday November 16, 2017 @05:41AM (#15711631)
"Recomputing the AlphaGo Zero weights will take about 1700 years on commodity hardware"

Oh well...

As much as the hardware advances are important, maybe there will be better ways to train networks with limited hardware

(training even a small CNN with limited resolution inputs on a CPU is slow)

Re:Leela Zero – Go engine with no human-provided knowledge by rurounijones on Thursday November 16, 2017 @06:17AM (#15711808)
Keep reading. It already has the capability for you to contribute using your local machine, folding@home style.
Re:Leela Zero – Go engine with no human-provided knowledge by margorczynski on Thursday November 16, 2017 @11:07AM (#15713755)
Didn't the Zero version from Google train on something like 10 TPU's? Or is it only inference?

Seems just the CPU or even GPU aren't really suited for this kind of task

Re:Leela Zero – Go engine with no human-provided knowledge by rockdoe on Thursday November 16, 2017 @12:15PM (#15714436)
"Train" as in do the gradient computation: 64 GPU.

Generate the data to do that in the first place: 2000 TPU, each about 45 TFLOPS (or 2-4 x a GTX 1080 Ti).

Re:Leela Zero – Go engine with no human-provided knowledge by NhanH on Thursday November 16, 2017 @11:17AM (#15713827)
AlphaGo Zero trained on 2000 TPUs