Some new details surrounding Nvidia's GP100 Pascal-based GPU have been uncovered this week, revealing that it is targeting some pretty high-performance numbers with 4 TFLOPS of double precision compute performance and three times that in single precision performance. This information dates back to a presentation given by Spanish university professor, Manuel Ujaldon in June last year.
These performance targets, if accurate, signal that Nvidia might be focussing once again on getting a 1:3 single to double precision performance ratio, which Nvidia stepped away from with Maxwell, which focussed more on single precision.
As Videocardz points out, this also gives us a hint as to how many CUDA cores the GP100 (‘Big Pascal) GPU will have. Under the assumption that Nvidia will keep the core clock at 1000 MHz, then this GPU could have 6144 CUDA cores, which is twice as much as GM200. This level of performance would likely be reserved for a future professional GPU or a TITAN series card.
Obviously, we don't know how accurate the original presentation was or if it would still remain accurate today almost a year later so this is mostly speculation. Either way, we should learn more in the coming months as we get closer to the official launch of Pascal.
KitGuru Says: A lot of people are waiting on Nvidia's Pascal launch this year so hearing about potential performance is pretty exciting. Will any of you guys be upgrading once Pascal hits?