| JEMC said: My comment was in regards to your density and performance increase guesses, not the waffer costs. After all, the article claims that Nvidia will restructure the SM, which should influence both those parameters. Given what's stated in the article, Mr. leather jacket already went to TMSC to secure 3nm wafers so, unless Nvidia has decided to split their chips between both TSMC and Samsung (for example going with the former for the bigger and more expensive parts, leaving the rest for Samsung), we have conflicting rumors. Which one to believe? |
That's all from TSMC. The performance increase at isometric power means the expected clock speed increases, so it's just basically what Ada at 3 nm would be.
As for the second point, it wouldn't be the first time Nvidia has done it. HPC Ampere was made on TSMC's N7 while GPU Ampere got Samsung's 8LPP. The end result was that AFAIK, the Ampere GPU parts have rather worse perf per watt than HPC Ampere.
Samsung's 3GAE has gate and metal pitches much closer to TSMC's 5N than 3N, so it would be a peculiar choice for an upgrade overall, definitely one you wouldn't use for HPC if you could.
