If a leaked presentation slide from CUDA Fellow Manuel Ujaldon is to be believed, NVIDIA’s upcoming GP100 GPU based on the company’s next-gen “Pascal” architecture will deliver at least 12 TFLOPS single precision and 4 TFLOPS double precision performance.
As you can see in the slide above, the Pascal GPU will bring an almost three-fold jump in DP performance compared to the Kepler based GK110 GPU. It is more or less the same story with the single precision performance numbers as well. If these numbers are indeed accurate, it is likely that the GP100 GPU may have to utilize 6144 shader units clocked at around 1000MHz. That’s almost twice the number of shader units the GM200 chip offers.
The slide also points at 1 TB/s memory bandwidth, confirming that the GPU will utilize stacked HBM2 memory.