davidcpage / cifar10-fast

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Not so fast

brycexu opened this issue · comments

Hi,

I ran your code on a single GPU, which is NVIDIA GeForce GTX 1080Ti, but training 24 epochs took me nearly 4 minutes.

I did not change anything in your code except two paths.

Is that we have to configure something before we run your model ?

Thank you.

A V100 vs the 1080ti will make a big difference I believe, because of the half precision stuff

The GPU makes a huge difference. For example, the 1080 Ti actually has worse GFLOPS with half precision compared to single precision (177 vs 10,339 GFLOPS). I think the reasoning was that NVIDIA would still be able to sell data center GPUs.