JEMC said: Consumer Pascal will run quite a bit faster than Maxwell parts. Nvidia (and AMD) Pro line of products always run at lower speed than consumer variants, yet the P100 runs at 1328/1480 MHz. The stock frequencies of Maxwell's 980Ti are 1000/1075MHz and the 980 ones are 1126/1216MHz. It's not too far fetched to think that the GTX 1080 will achieve a boost clock of 1500MHz at stock speeds. With that and the advantages that the new architecture could bring (there must be a good reason to shift the focus to more SMs rather than more cores per SM), I keep saying that the 1080/1070 will bring a nice upgrade over not only the 980/970 but also the 980Ti. Guess I'm going to be the optimistic one here. |
Nvidia's professional parts don't run THAT much slower as of recent. A GTX Titan X runs at a base clock of 1000MHz. The Quadro M6000 runs at 988MHz and the Tesla M60 Accelerator runs at 948MHz. They only run 12MHz and 52MHz slower respectively base clock wise all on a 250 watt TDP. The difference is practically negligible! The GP100 is already eating up it's overclocking headroom by running at a 300 watt TDP for comparison ...
Boost clocks are not what we should be focusing on, it should be base clocks since they guarantee a baseline ...
While Pascal did change the SM arrangement, you need twice as many work items as you did before. Nvidia is practically trading in IPC/ILP for more thread level parallelism ...
I see no reason to believe that we're going to get a higher IPC with Pascal when it has a lower amount of shared memory per thread ...