During his GTC 2021 keynote, Nvidia CEO Jensen Huang revealed a roadmap for its upcoming GPUs, CPUs, and DPUs. We've already covered the Grace CPU, which should arrive in 2023, but the more interesting bit for gamers is the GPU. While Nvidia listed the GPU as "Ampere Next," we expect the GPU codename will actually be Hopper. That means Nvidia's next CPU and GPU combined will be Grace Hopper, after the computer scientist who pioneered many aspects of computer programming languages.
We don't have any concrete details on Hopper / Ampere Next yet, though it's safe to assume it will be faster and better than Ampere. Perhaps more importantly, it won't arrive until 2022, so anyone hoping for a sequel to Ampere this year will have to wait a bit longer. Which is perhaps all for the best, considering that it's basically impossible to buy any of the best graphics cards right now — unless you're willing to pay two to three times the official prices, anyway.
What will Hopper / Ampere Next entail, though? Current Ampere GPUs are manufactured on both TSMC N7 (GA100) and Samsung 8N (all of the consumer models — GA102, GA104, and GA106). That split could continue, but almost certainly the data center and HPC solution will come via TSMC, probably on the N5 process. That change should allow for even more GPU cores and transistors. The consumer side could potentially switch to TSMC as well, depending on how quickly TSMC can ramp up its wafer starts.
Based on past releases, we expect the Ampere Next / Hopper GPUs to launch with RTX 4000 series branding, and the top model (ie, RTX 4090) would likely outperform the current offerings by at least 25–30 percent, in situations that are GPU limited. We'll probably see fourth-gen Tensor cores and third-gen RT cores, each boosting performance over the current solutions.
And unless things have dramatically improved by 2022 [Ed: Dreams are free!], they'll all still sell out immediately after launch, at least for a few months.
But there's still hope for the follow-up to Ampere Next / Hopper. That is currently listed as "Ampere Next Next" in the roadmap, and we don't have a guess as to the future codename (Babbage, Lovelace, Euclid... feel free to guess your favorite scientist). We do know that Nvidia expects it to ship in 2024, which means it will probably utilize TSMC's N3 process or something similar, cramming even more GPU goodness into the package. By that point, we might even have consumer GPUs with more than 100 billion transistors.