Nvidia Pascal GPU Specifications Predictions: Six SMs per GPC
While Nvidia is not opening up about the details, and Pascal GP100 is going to take more than a few months to be out in thee open, predictions are what we will be relying on for the Nvidia Pascal GPU specifications.
VideoCardz has done the hard work for us on this. Assuming a couple of things like what could be the architecture’s positioning, that the whole thing is based on a similar solution to Maxwell and Kepler in terms of GPC splitting.
So firstly, they are suggesting that in comparison to three Streaming Multiprocessors per Graphics Processing Cluster of Kepler and four that of Maxwell, the Pascal GPU specifications include six Streaming Multiprocessors per Graphics Processing Cluster.
However, Kepler has 192 CUDA/SM and Maxwell has 128 which compares to a very low 64 CUDA/SM of Pascal. CUDA/GPC also varies proportionately at 576 and 512 for the existing architectures and 384 for Pascal.
For Pascal GP100, in terms of enthusiast Nvidia GPUs, the Die Size is 610 mm2 with 15.3 billion transistors resulting in 25.08 million transistors/mm2.
They also talk of Pascal GP104 where the die size is approximately between 350-400 mm2 (most probably around 400) and approximately 10.2 billion transistors with 4 GPCs, 40 SMs and 2560 CUDAs.
Moving on to Mid-range Pascal GP106; it is expected to have a die size between 190 to 215, with approximately 5.4 billion transistors, 2 GPCs, 20 SMs and 1280 CUDAs. Suggesting a smaller bump over GM206.
Lastly, we have the entry level Pascal GP107 where a 120-150 mm2 die size is predicted with 3.5 billion transistors, 1 GPC, 10 SMs and 640 CUDAs.
While that is all we have for Nvidia Pascal GPU specifications predictions right now, we will be more informed in the coming days and weeks.