Over at WCCFTech a news has surfaced, where interesting details on NVIDIA's upcoming Pascal GPUs, codenamed GP100, were outlined. It looks like NVIDIA will be implementing HBM2 memory with up to four stacks, which will be up to 8-Hi. Apart from that a 4096-bit memory interface is planned.

If the report is correct, there will be cards with different HBM2 implementations. The consumer models (GeForce) will be equipped with 4-Hi HBM2 stacks running at 1 GHz, whereas the 8-Hi stacks will only be available to the professional market (Quadro and Tesla), which demands as much video memory as it can get.

NVIDIA's Pascal GPU are expected to be launched during next year. By then TSMC should have the 14/16nm FinFET process ready. So far it can be expected that NVIDIA will make TSMC manufacture their GP100 GPU using 16nm process technology. Therefore there will not only be a significant performance improvement thanks to the new memory and possible other changes to the architecture. Switching from 28nm fabrication to 16nm process technology is a huge step and it means that energy efficiency of the GPU will be on a completely different level. Comparing the current 28HPM process with the upcoming 16FF+ up to 65% higher clock speeds can be achieved whereas 70 percent less power is needed when running the same frequencies. If the complexity of NVIDIA GPUs is not going to increase drastically, we could expect to see higher clock speeds and maybe Pascal chips could even do 2 GHz on air.

We expect that Pascal is going to be a huge advancement over Maxwell. The combination of HBM2 and 16nm process technology sounds very promising and NVIDIA graphics cards would greatly benefit from highger memory bandwidth. Apart from that the new process will definitely allow of higher clock speeds. If we were to guess we would say, that there is a good chance that NVIDIA Pascal GPUs might end up more than 50 percent faster than Maxwell.




Source: WCCFTech