NVIDIA has announced their new Tesla P100 supercomputer chip that can be plugged in using the standard PCIe bus for those without an NVLink CPU.
The company unveiled the NVIDIA P100 in April during their GPU Tech Conference. It is built through the 16nm FinFET method which was also used to make the consumer grade GTX 1080 and GTX 1070 Pascal cards.
NVIDIA's new accelerator chip is a huge upgrade from their previous K/M series that used the 28nm Kepler and Maxwell GPU architectures. It is also packed with more features and more powerful specs for servers and supercomputers.
For starters, the new card has 3585 stream processors. NVIDIA's Tesla P100 PCIe GPU can dish out 4.7 TFLOPs for 64-bit Double Precision crunching and an astounding 9.3 TFLOPs for 32-bit Single Precision data crunching, Anandtech has learned.
Since it would not be using an NVLink CPU, it only has a TDP count of 250W which is lower compared to the Mezzanine's 300W. It has 16GB of VRAM and is passively cooled.
Intel does not currently support the NVLink architecture and it is one of the reasons why NVIDIA has launched the PCIe Tesla P100 GPUs, The Register reported. Those who want to build their servers can now use the new GPU with their x86 computers.
The performance decrease from the NVLink Mezzanine version of the card is not that much at all. The Double Precision performance of the non-PCIe variant can dish out 5.3 TFLOPs, 10.6 TFLOPs for Single precision and 21 TFLOPs for half-precision compared to 18.7 TFLOPs from the PCIe bus model.
Unlike the GTX 1080's GDDR5X, the Tesla P100 has 16GB of HBM2 stacked RAM which provides data bandwidth of up to 720 GB/s. There is also cheaper option that has 12GB of HBM2 RAM which has a significantly lower bandwidth of 540 GB/s.
IBM, HP, Dell and Cray will also be selling the supercomputer cards besides NVIDIA. The final price of the card hasn't been revealed yet but it could be around $4000 for the 12GB variant.
NVIDIA will be rolling out the Tesla P100 PCIe bus variant in the fourth quarter of 2016. Reseller's price will also vary from the cards that will be sold by NVIDIA themselves.