[cobookman]

If all you need is reduced precision, then it’s fair to compare the two. I’d also assume memory bandwidth matters just as much as TFLOPS for ml workloads.