Nvidia has released a whitepaper on the Pascal GP100 GPU architecture, which will be used in the recently announced Tesla P100 as well as the rumoured high-end GeForce graphics cards.
The 45-page Pascal whitepaper states that the Pascal GP100 GPU will be coming with with 1,920 FP64 CUDA cores and HBM2. In addition to that, Nvidia has mentioned that its upcoming high-end GPU will also be equipped with 4,096KB of L2 cache – about 1,000KB more than the Maxwell GM200 GPU.
There is, however, a slightly worrying fact mentioned inside the Pascal GP100 whitepaper: “Compute Preemption”. According to Nvidia, Compute Preemption “prevents long-running applications from monopolising the system or time out”. This could mean that Nvidia will not look into supporting any form of Asynchronous Compute tasks on its upcoming Pascal GPUs.
Unfortunately, like all other top of the line GPUs by Nvidia, some of the features of the full Pascal GP100 GPU may not see the light of day when it comes to consumer-grade products. However, a Pascal-based GTX Titan X refresh may be powered by an unaltered, raw GP100 GPU instead.