Re:Apple M1 Max Geekbench Score
by throwawaylinux on Wednesday October 20, 2021 @10:15PM
(#28939492)
Wow that's stupid, I didn't look that closely. So it's really a 5x perf/watt improvement. I assume it will be the same deal for this, around 5-6x perf/watt improvement. Which does make more sense, FP16 should already be pretty well optimized on GPUs today so 30x would be a huge stretch or else require specific fixed function units.
[ Parent
| on HN ]
it's an odd coincidence (there's no reason this number would be related, there's no idle power factor here or anything) but 5x also happens to be about the expected gain from NVIDIA's tensor core implementation in real-world code afaik. Sure they advertise a much higher number but that's a microbenchmark looking at just that specific bit of the code and not the program as a whole.
it's possible that the implication here is similar, that AMD does a tensor accelerator or something and they hit "30x" but you end up with similar speedups to NVIDIA's tensor accelerator implementation.
I've seen tensor cores really shining in... tensor operations. If your workload can be expressed in convolutions, and are matching the dimensions and batching needs of tensor cores, there's a world of wild performance out there...