Nvidia's Ampere architecture is barely out of the gate, especially when you factor in continued shortages of the new GeForce RTX 3080 and GeForce RTX 3090 — the theoretically top picks for the best graphics cards and GPU benchmarks hierarchy. The latest rumor from Digitimes (also discussed at TechPowerUp), however, is that the follow-up for Ampere is already on the way and could arrive as early as 2021. #MuchSalt #Speculation
To be clear, we've already mused about this ourselves with the Ampere launch. There's no question that TSMC's N7 process is superior to Samsung's 8N in nearly every important metric — except price and available production capacity. Besides N7, TSMC also has N7P, N7+, N6, and N5 nodes all in active production. In other words, there are a lot of options available for Nvidia to consider, any one of which should be better than Samsung 8N, provided TSMC and Nvidia can come to some sort of agreement on price.
Considering Nvidia CEO Jensen has talked about Nvidia's desire to diversify and not be reliant on any single supplier for chips, plus the shortages of RTX 30-series GPUs, it makes sense for Nvidia to think about shifting some future production over to TSMC. And of course, with RTX 3090 and 3080 being the two highest power single-GPU solutions Nvidia has ever shipped to consumers, finding a way to reduce power requirements and potentially boost performance would be great. However, there are many other factors that also come into play.
First, Nvidia would have to have new designs (relative to the current RTX 30-series GPUS) for TSMC. Samsung 8N is different from all of the TSMC options, so some additional porting work would be required. It shouldn't be too complex, relatively speaking, since Nvidia already has its A100 GPU shipping on TSMC N7, but porting a design to a new manufacturing node is never 'simple.'
Second, we have to assume that any GPUs manufactured on one of TSMC's advanced nodes would perform better than the current GA102/GA104 designs. We would expect improved power characteristics, potentially higher performance, and smaller chips from N7, and additional improvements should Nvidia opt for N7P, N7+, N6, or N5 — though the last three would be even more complex ports, considering they use EUV.
That leads us to the logical conclusion that if Nvidia does end up porting Ampere over to TSMC, it won't be for additional sources of the current RTX 30-series GPUs. Instead, it will be for a mid-cycle refresh similar to last year's 20-Series Super models (i.e., RTX 2080 Super, RTX 2070 Super and RTX 2060 Super). Otherwise, selling two rather different variants of the same chip would lead to problems. Even if the TSMC versions were simply lower power, that's a big enough change that someone would inevitably file a class-action lawsuit. Releasing as RTX 30-series Super avoids that possibility.
There's another possibility, though. The rumor mill suggests Nvidia's post-Ampere architecture might be called Hopper, and this could all just be Nvidia prepping its next-generation GPUs. Hopper is anticipated to use TSMC N5, and normally we'd expect it to arrive 18-24 months after Ampere. Ampere was at least somewhat delayed thanks to COVID-19, however (we originally expected a spring or early summer 2020 launch), so a late 2021 launch isn't out of the question. Likely a big factor will be just how competitive AMD's Big Navi ends up being when it's revealed in a few weeks. Should AMD end up matching RTX 3080, maybe even RTX 3090, Nvidia will almost certainly fast-track a TSMC-based successor.