EricHiggin said:
I'd file it under sorta maybe but not really. I think it was the most recent presentation where Nvidia showed how much more performance the upcoming cards were going to have vs past cards, and they used like FP8 for the upcoming cards, FP16 for the present cards, and FP32 for the prior gen. Which of course made the performance gains look monstrous, instead of the truth with each being shown with the same measurement. You don't do that if you're performance/efficiency is good, and that hasn't been the case for a few gen's now for Nvidia. Radeon cards aren't exactly sipping power either though. |
That's true, but Huang Maths has been a staple of Nvidia slides for a long time, even when the performance improvements were actually good.
AMD is doing the same thing (25x more "AI efficiency" in five years or something).
