Hacker News Reader

Cost per GFLOPS: https://en.wikipedia.org/wiki/FLOPS#Hardware_costs

Probably the clearest chart that explains why machine learning is blowing up now, despite the fact that the same neural network architectures existed since 1990s (e.g., LeNet-5).

It's because cost dropped from $47,000 per GFLOPS to $0.02!!!

View on HN

That $0.02 value is entirely bogus. It is based in the published 30 "teraops" for Xavier, which it uses as teraflops and assumes is comparable to the other flops values. ($699/30 "TFLOPS" = $0.02 $/"GFLOP"). "Teraops" counts int8 performance on the GPU plus the random vision and CNN ASICs in Xavier - it is definitely not equivalent to normal fp32 flops.

It looks like the GPU in Xavier can do 1.4 TFLOPS single precision, so the correct value is $0.49.

Reply on news.ycombinator.com