New Google result on DAWNBench

Thought this would be relevant to the ongoing Imagenet training project–Google just achieved a new mark of 7:28:30 for training time on Imagenet, beating their own previous record by 5 hours. (https://dawn.cs.stanford.edu/benchmark/#imagenet-train-time).

They were able to do so using an AmoebaNet architecture, which was developing using evolutionary search as described in this paper: https://arxiv.org/abs/1802.01548

2 Likes

Yeah I’m not too happy about this… :wink:

4 Likes

Hmmm…How fast do you think we can train imagenet to 93% with their tpu and our cutting edge training methods? :slight_smile:

And they just hit 30 minutes with half of a TPU pod (32 TPUs) https://twitter.com/JeffDean/status/987118479471005698