For the past six months, the consumer GPU market has grappled with three basic questions: When would GPUs be affordable, when would Nvidia launch a refresh to its popular Pascal architecture, and when would AMD launch new parts to hopefully better compete with Nvidia? So far this year, we’ve gotten lower GPU prices , an Nvidia launch event that might or might not happen this fall (rumors on timing and plans and possible inventory hold-over are running rampant), and from AMD’s corner… not much at all. Publicly, the only thing the company has said is that it plans to launch a 7nm Vega refresh for the machine learning market later this year.
But there’s a rumor from Chiphell user WJM47196, who has a fairly good track record on these sorts of announcements, that’s making the rounds on what AMD might be planning for this year — and it’ll probably raise some eyebrows.
First off, there’s a new Polaris family supposedly being prepped for a Q4 2018 introduction. Built on the same 12nm process improvement as AMD’s second-generation Ryzen CPUs, it would offer a 15 percent performance improvement over current cards.
There are only two ways for AMD to deliver this kind of performance improvement, and only one of them makes sense. To improve GPU performance you can make a GPU wider or smaller while keeping the architecture the same, or you can increase the GPU’s clock speed. With architectural changes unlikely, this suggests that AMD would instead simply try to ratchet clocks higher. And who knows? Given that Polaris was the first 14nm discrete GPU GlobalFoundries had ever built, it’s possible that AMD found some low-hanging fruit that would allow it to hit higher clocks. The larger problem is that while a 15 percent jump would give AMD much stronger footing against Nvidia today, it might not compare well against any next-generation hardware Nvidia launches — at least not without price cuts.
Next up, Navi. There’s a rather confused suggestion that Navi will be both a mainstream and high-end part arriving sometime in the 2019 timeframe, and that it will debut in the budget segment first before eventually launching as a high-end, HBM2 equipped part sometime “much later.” The suggested time frame is:
Q4 2018: Polaris 30 (performance up 15 percent).
H1 2019: Navi 10 (budget part, and timing on this introduction is unclear, with additional reference to a Q1 release)
H2 2020: A new, high-end Navi part, as a “true” successor to Vega
This is rather muddled, and I’m not sure at all how much weight I’d put on it. AMD has been widely reported to be working on Navi for Sony’s PlayStation 5 and we know the company has often aligned its console and non-console launches. The idea that the work AMD is doing for Sony could lead to a budget GPU built on an early 7nm process as a pipe-cleaner? Not crazy. Similarly, the idea that AMD would port Polaris to 12nm and spin it for a bit of additional headroom isn’t nuts, either.
AMD’s larger Polaris GPU (pictured above)
If we had to make a bet, here’s where we’d land. The Navi 10 chip, if it exists, is a custom 7nm GPU built for a customer like Apple. This chip gives AMD a better foot into ultra-low-power markets, where its relatively power-hungry GCN architecture can use all the process node improvements it can get. The overhauled Polaris lineup, meanwhile, gives AMD a 15 percent performance kick in desktop, where it can be combined with price cuts to boost AMD’s overall position. This would be further strengthened if Nvidia focuses on refreshing its top-end GPUs first, as the company is likely to do. It wouldn’t be unusual for the introduction of new midrange cards to lag the launch of a GTX 1180/1170 for some months.
The idea that Navi is a single GPU design that stretches across both GDDR6 and HBM2 is a little odd, and likely reflects the fact that Navi is being used to refer to a new architecture rather than a single card. There’s precious little reason why AMD would attempt to take a single GPU across both memory standards when the costs of doing so are so high — the entire GPU memory subsystem and memory controller both have to be rearchitected when moving from GDDR6 to HBM2 or vice-versa. This may not be the only reason why Nvidia has kept its non-Pascal GPUs almost entirely in the professional space, but it’s undoubtedly part of it. But the idea that it could take AMD until Q3/Q4 2020 to replace Vega had better be wrong.
It took AMD ~26 months to move from Fury X to Vega. At the time, this was the longest high-end refresh cycle in history, though Nvidia has actually taken back that record. If the GTX 1180 launches in August, it’ll have taken Nvidia 27 months to pull off a high-end refresh. But Vega is already almost a year old, it didn’t cleanly put AMD back in competition with Nvidia at the highest level, it isn’t going to stand up well against the GTX 1180 or hypothetical 1180 Ti without price cuts to reposition it (not unless Nvidia simply raises its own prices, which, hey, it could do), and the idea of leaning on it as a high-end solution for another two years after an Nvidia refresh is a poor one. It would effectively imply that AMD, which is already operating a GPU cycle behind Nvidia, would be willing to sit out at the high end entirely for another two-year period.
But as we’ve said, this is a rumor — and a rather disjointed rumor at that. Take it with a small mountain of the white stuff.
Now read: PCMag’s Best Graphics Cards of 2018