05-16-2016, 01:11 PM
(This post was last modified: 05-16-2016, 01:11 PM by Chick3nman.)
(05-16-2016, 12:44 PM)Rolf Wrote: I've read elsewhere that watercooled 1080s may have core clocks of 2.5 GHz and according to that source it may be enough to beat GP100 performance-wise at some tasks.
WCCFTech is notorious for spreading rumors and BS though. I'd take everything with a grain of salt from them. They are banned as a source on several subreddits due to their wild claims with 0 evidence to back them up.
Even in that article, some of the things they are saying don't make sense. Of course the 1080 has higher single precision than the P100 when overclocked, but even though they are both pascal, i wouldnt consider it a fair comparison. The P100 is designed for a different purpose and actually has a different die layout as well with only 1 block of 64 CUDA cores per SM instead of 2 like the 1080. It also was focused on FP16 performance and memory bandwidth for deep learning simulations, and has a measurable double precision speed that i doubt the 1080 has(assuming nvidia went the same route as they did with maxwell and basically cut double out for more single precision).
This source has a bit of the same information though not quite as hyped up.
http://vrworld.com/2016/05/10/pascal-sec...-gtx-1080/
No claims of 2.4GHz on air cooling, though they are saying that the 2.5GHz reported overclock may be reached via Water cooling, which makes me doubt the validity of 2.4GHz on just air cooling(unless that was run up against the thermal limits pretty closely). Also sounds like the AIB cards are going to be using a good bit of power over the stock 180W to run at those speeds according to this, with the 225W limit of the single 8pin+pci-e being too little for anything over 2.2GHz.