Training Speed on AWS p2.xlarge

performance
pytorch

(Benjamin Breton) #1

Hi, I am training an english to french model on the CommonCrawl dataset with the pytorch implementation and the default options.

I am training it on 2 machines:

  • My Laptop with a i7, 32gb of Ram and a GTX 1050Ti (4Gb of Vram) : training speed ~2000 tok/sec
  • AWS p2.xlarge: Xeon, 60GB of Ram, Tesla K80 (12GB of Vram): training speed ~2000 tok/sec

It is very strange (especially considering the price of the AWS machine). Does anybody have an explanation?


(David Landan) #2

The K80 is not particularly fast. There are now AWS instances with P100 cards; these are much faster. I did some benchmarking with the K80-powered p-series in the past, and your results are in line with my experiences.