Rumor – NVIDIA GP100 Pascal GPU To Have 4 TFLOPS of Compute Peformance

Gamers and enthusiasts are getting excited about NVIDIA’s upcoming next-generation flagship GP100 GPU based on the brand new ‘Pascal’ GPU architecture. The folks over on the 3DCenter.org forums found a slide by NVIDIA CUDA Fellow Manuel Ujaldon that shows rather impressive double-precision floating-point (DPFP) performance! The slide shows 4 TFLOP/s when it comes to double precision compute performance, which is more than a triple the 1.31 TFLOP/s found on an NVIDIA Tesla K20 card that uses the NVIDIA GK110 ‘Kepler GPU.

We dug a little deeper and found that the slide is from “New hardware features in Kepler, SMX and Tesla K40” presentation that took place in April 2014. This is rather old information and it’s very likely that these nice round figures shown for Pascal in the slide have changed over the past two years. The slides mentions that a Pascal with stacked memory, but remember this slide is from 2014 and it’s likely HMB2.