vivster said:
A lot of that research went to Volta and the most expensive part of it was the shrink which is always very expensive. Increasing yields and optimizing manufacturing techniques always takes a lot of time and money. The architectural move from Kepler to Maxwell was a gigantic leap and showed what kind of performance you still could get in the old 28nm process. The change from Maxwell to Pascal is basically nothing. Volta will be the next big leap. It will not have the same performance jump but it will lay the groundwork for optimizing the 16nm process and finally feature HBM2 for mainstream cards, one of the most important components when it comes to 4k and above. It will most likely feature around the same clocks so it's basically forced to innovate in chip design to gain performance over Pascal. Compare it to Intel's Tick Tock model where they basically change nothing at the shrink but will still get better performance and greater efficiency through clock and packing density. |
Interesting. So the CEO kind of lied at the presentation then if the research actually went to Volta lol.
I didn't know the shrink was so expensive to Nvidia research-wise, because that would be on the foundry. And the foundry then recoups its investments in the process shrink by charging Nvidia for each GPU manufactured, but that would never be counted as R&D costs for Nvidia.
And Pascal was extremely hyped before release, not only because of the process shrink. So I don't get how all of a sudden Volta is the one where architecture miracles are supposed to be made. But I hope you're right.
