When Nvidia demoed Pascal at GDC 2016 last month, many readers were a bit unhappy that the company didn’t give more details on what its upcoming consumer cards would look like. The latest rumors are that Nvidia may have stopped production on its GM204 products like the GTX 980 Ti, 980, and 970 in order to quickly replace those products with a GP104 derivative.
GP104 is the name for Nvidia’s next-generation consumer-level Pascal card, and it’ll debut long before the HPC-oriented GP100 does. The full scientific version of Pascal isn’t scheduled to launch before the tail end of this year, whereas we expect to see these new consumer cards launching as early as June. This would put AMD and Nvidia on similar time frames (AMD has yet to reveal its launch dates, but we expect cards in June or July).
Nvidia released the full whitepaper on Pascal GP100 last week, and while we can’t draw many conclusions at this juncture, there are a few things we can safely assume will be common to the two architectures. First, Pascal supports improved compute preemption compared with Maxwell. From Nvidia’s whitepaper:
Compute Preemption is another important new hardware and software feature added to GP100 that allows compute tasks to be preempted at instruction-level granularity, rather than thread block granularity as in prior Maxwell and Kepler GPU architectures. Compute Preemption prevents long-running applications from either monopolizing the system (preventing other applications from running) or timing out.
This suggests that Pascal still won’t support asynchronous compute workloads in the same fashion AMD’s GCN hardware does, though the verdict is still out on whether this capability will play a significant role in shaping performance in a majority of DirectX 12 titles. It also suggests Pascal will see a smaller performance penalty with async compute enabled than Maxwell did, since it can interleave compute workloads more effectively than its predecessor.
The image above shows the structure of a Pascal GP100 SM unit. Pascal is heavily focused on double-precision compute, which means the GP104 variant of the chip will probably remove these units altogether, then use the space savings to pack in more single-precision FP32 cores. Expect limited double precision support, just as we saw with GM204 and GM200.
One significant difference would be the way Nvidia hits its targets. Kepler and Maxwell both relied on two different dies — the highest end cards were anchored by GK110 / GM200, while the high-end segment was anchored by GK104 / GM204. These rumors suggest that Nvidia will replace the GTX 980 Ti with the full iteration of GP104, while the future GTX *80 and *70 cards will be further cut down from this base model.
As for the bigger picture, I’d expect AMD and Nvidia to both launch GPUs with 8GB of RAM and possibly some 4GB cards depending on price brackets and targets from the two companies. Polaris and Pascal are likely to drop within weeks of each other, so I wouldn’t run out and buy a card from one company before we’ve had a chance to put them head to head.