What does Nvidia have planned for the next round of graphics cards? The Turing architecture and 20-series first launched with the RTX 2080 Ti in September 2018, and this year brought us the Super refresh. It doesn't take a GPU architect to know that Nvidia is already well into the design stages for the next generation post-Turing graphics cards, which will most likely be called Ampere.
To be clear, Nvidia hasn't officially told anyone what the next-generation GPU will be called, though it mentioned Ampere in the past (before Turing ever launched). And it's certainly not going to leak specs—anyone claiming to have specifications for the RTX 3080 is just making stuff up. Which is par for the course, as it happened with the RTX 2080 and 2080 Ti, and the GTX 1080 Ti before that, and so on. (For the record, nearly every guess ended up being wrong.) But even if we don't know the exact specs, looking at the underlying tech—in this case, TSMC's and Samsung's 7nm process technologies—we can get a good sense of what to expect.
Let's start at the top, with the GPU cores. Nvidia GPUs build on each preceding generation. Turing in many ways is like Pascal, with some extra stuff added into the mix. Pascal is like Maxwell, with extra stuff. Pascal also marked a pretty major transition from 28nm lithography to 16nm FinFET. And before Maxwell there were two generations of 28nm Kepler GPUs. This next transition is going to be most like the Maxwell to Pascal move, though, because it's not just about updating and tuning the architecture.
Back in 2014, when it became clear that the benefits of TSMC's 20nm lithography weren't that great, Nvidia reworked many aspects of its graphics architecture and released Maxwell. Still using the same 28nm tech that was behind the GTX 600 and 700 series parts, the GTX 980 and 970 were nevertheless amazing cards at the time—even now, the GTX 970 still does okay, trading blows with the likes of the GTX 1060 3GB and the GTX 1650. The new $330 part came pretty close to matching the performance of the outgoing $700 780 Ti. Then two years later, Pascal's $380 GTX 1070 would once again basically match the performance of the outgoing $650 GTX 980 Ti.
The next GeForce series—RTX 3080, Ampere, or whatever they end up being called—will be a major transition for Nvidia's GPUs. I don't expect a massive change in the underlying architecture, but the shift from 12nm to 7nm should pave the way for some big improvements. I also expect Nvidia to be a bit less ambitious on the pricing, as the RTX 20-series by all accounts didn't sell as many units as Nvidia hoped. The lack of interest from the cryptocurrency miners probably had a lot to do with that.
Looking to AMD's 7nm parts, we can get a pretty good idea of how the 7nm process scales. The Vega 20 chip in the Radeon VII measures 331mm^2, compared to 486mm^2 on the Vega 10 GPUs. That's about 30 percent smaller, while it also packs in 6 percent more transistors. What does that mean for Nvidia? It means a die shrink of the TU106 used in the RTX 2060 and 2070 cards could end up being the same size as the TU116 used in the GTX 1660 cards. It would also use less power, or alternatively, it would use the same power and run at higher clockspeeds.
More likely than a straight die shrink is that Nvidia will pack more 'stuff' into the next generation Ampere GPUs—more RT cores for ray tracing, more CUDA cores for general purpose graphics, and more Tensor cores for machine learning. There will probably be some new tech to improve performance as well, because Nvidia tends to do that every generation.
Best guess right now: Nvidia will put 20-25 percent more cores in each level of GPU hardware (eg, so the 3080 will be sporting up to 3840 or maybe 4096 cores next round). The 7nm process will allow Nvidia to clock those cores 20-25 percent higher as well, which means for the first time Nvidia could have GPUs consistently running at more than 2.0GHz. And finally, even with the added cores and features, die sizes will end up smaller than the current Turing equivalents. We could be looking at a generational performance improvement of 40-50 percent or more for RTX 3080, with a chip that's still 15 percent smaller than the RTX 2080.
If that seems like too much to hope for, look at AMD's GPUs again. The Navi 10 GPU in the RX 5700 XT is a 251mm^2 part that uses 225W, and it easily outperforms the RX Vega 64—a 486mm^2 part that uses 295W. Some of that is architecture, but plenty of it comes from TSMC's 7nm lithography. Nvidia is already able to beat AMD's best performing GPUs, the RX 5700 XT and Radeon VII, with existing cards: the RTX 2080, 2080 Super, and 2080 Ti are unequivocally faster, and the 2070 Super leads in most games as well. Nvidia's next generation GPUs aren't going to get any slower.
How soon could Nvidia's 7nm Ampere GPUs arrive? TSMC has been manufacturing 7nm chips since 2018, when the Apple A12 launched in quantity in September. AMD first started selling 7nm in late 2018, and it has been selling lots of 7nm CPUs and GPUs since this July. Nvidia is likely to announce and launch its first 7nm parts by spring 2020, though it could be a few months before or after that target.
If you're still sitting on a GTX 10-series or similar hardware, or maybe even the GTX 900-series, you may finally have a reason to upgrade next year. And wouldn't it just be perfectly timed if Nvidia launched RTX 30-series parts with substantially better ray tracing performance, just in time for Cyberpunk 2077? That's my bet. Check back in six months and we'll see how I did.