NVIDIA has developed the NVIDIA Tesla P100 cards, designed for Data Centers with great computing power and that make use of HBM2 memories.

We continue without signs about the AMD Vega graphics cards and while NVIDIA has continued to launch products, such as the GTX 1080 Ti and the Titan Xp and now the professional graphics card NVIDIA Tesla P100, developed for Data Centers, is added. This graphics card connects through a PCIe and will come in two versions, one with 16GB HBM2 and another with 12GB HBM2, in addition to a variation of the NVLink interface that has been optimized for systems with 16GB HBM2. The catalog is therefore with three versions of the NVIDIA Tesla P100 that offers different solutions to cover all kinds of needs.

This Tesla P100 graphics card uses a Pascal P100 GPU in its full version, or what is the same, with the 3840 CUDA Cores. This GPU has 240TMU and is accompanied by 16GB HBM2 and no less than eight 512-bit memory controllers, thus generating 4096bits of memory interface for a total bandwidth of 720GB / s. The GPU is made with the 16nm FinFet architecture and this should offer an improvement in the overall performance of this graphics card with significant energy savings.

The NVIDIA Tesla P100 is able to offer us a computing power in simple performance of 9.3TFLOPS, while the NVLink model offers 10.6TFLOPS and in double precision it offers 4.7TFLOPS, which for the NVLink model is 5.3TFLOPS. As we have mentioned, the 16GB MHB2 model has a memory bandwidth of 70GB / s, while the 12GB HBM2 model offers a bandwidth of 540GB / s. Both have a passive heat cooling system, although they have a TDP of 250W, which is because server systems are usually very well cooled. They have not revealed the price, nor does it need to be done, since they are components designed for very specific systems and are not for the general public.

Source: NVIDIA