Nvidia Confirms: Pascal Flagship Graphic Card will Have 16GB HBM2 Memory and Boast 1TB/s Bandwidth at Launch
The Japanese version of the Nvidia GTC event (Via VRWorld) was recently held and the company confirmed some previously leaked details about the upcoming Pascal GPUs. The next generation GP100 flagship will indeed boast HBM2 technology but will be limited to 16GB HBM2 with around 1TB/s of bandwidth at launch. The Tesla segment of Nvidia GPUs will also recieve NV link which will bypass the bottleneck of the PCI-e bus by using a custom interconnect that is roughly 5 times faster.
Nvidia Pascal GPU will boast 16GB HBM with 1TB/s bandwidth - will only get 32GB HBM2 once the tech matures
It was previously believed that the professional variants could have upto 32 GB HBM2 memory, whoever, Nvidia has further clarified that this amount can be increased to 32GB HBM2 only depending on the state of the memory technology in 2016 (and depending on how much progress SK Hynix makes). This means that all Pascal products at the time of launch will be equipped with a maximum of 16GB HBM2 memory. Ofcourse this means that on the professional side of things, once ECC (Error Correcting Code) is added, the actual capacity should decrease.
The flagship chip of Pascal architecture is ofcourse the GP100. It will feature DirectX 12 feature level 12_1 or above and is built on the 16nm FinFET+ TSMC process. The chip allegedly has 17 Billion transistors – which is more than twice that of the GM200 chip. Zauba records indicate that it taped out in June 2015, so we can expect it in the first half of 2016. The Pascal Test Vehicle acutally had 4 GB of HBM1 memory, but as we now know, the production graphic cards will contain 16GB of HBM2 memory. The card will have full support for FP16, FP32 and FP64 tasks and will come in two form factors: namely that of the PCI-e form and the Mezzanine High Bandwidth form to be used with NV Link.
Nvidia has also stated that internally the GPU will surpass speeds of 2TB/s - something which means that the next generation product will truly be a distinguished from its predecessors. It is also worth nothing that Nvidia has actually skipped Maxwell on the dual GPU professional side and will skip directly to Pascal. Pascal architecture is thought to fix all the faults of the Maxwell 2.0 architecture - and finally perfect full DirectX 12 compatibility at the 16nm FinFET process.
Unfortunately for us, Nvidia has started scrambling their codenames (for obvious reasons) of GPUs shipped for testing - so we can no longer pinpoint the exact graphic card models. However, the pricing and timings of the recent Zauba shipping entries indicates that the Pascal GPU is nearing the end of its testing and will soon shift to CS sampling - which is the finished, final product. The exact date of launch remains unknown - and we have strong reason to believe that the delay will depend entirely upon sampling and yield of HBM2. Since 16nm FinFET+ is actually based on the 20nm backbone, and considering TSMC has had ample to time to mature the process, the node itself should have reasonable yield by early 2016. HBM2 sampling on the other hand, is one front where we currently don't have any conclusive reports on.