Here’s a really useful benchmark of cloud GPUs from Renaldas Zioma. It uses popular Convolutional Neural Network models to classify the performance of various cloud providers.
The benchmark was recently updated to include the new NVIDIA P6000 GPU. In Renaldas’ words, the “P6000 is … > x3 times faster than brick-and-mortar of clouds K80”. P100 benchmark coming soon
The author recently updated the benchmark to include the new Volta chips. It looks like he’s not using mixed precision training https://docs.nvidia.com/deeplearning/sdk/pdf/Training-Mixed-Precision-User-Guide.pdf yet – that should really speed things up. Would love to see someone get this going so we can see the results. Either way, this looks to be the fastest chip out there right now and the results are pretty insane.