[iaml]

GPT3 is so big it would take 355 years to train on a nvidia V100, so your example is also not really useful for comparison. It would be interesting to see some mid-sized nn benchmarks though.