NVIDIA Corporation introduced the graphics accelerator Tesla P100 in the version for PCIe-servers. The novelty is the expansion card with PCI Express 3.0 interface. Recall that in April, NVIDIA introduced accelerator computing Tesla P100 in the version for servers with NVLink. Both accelerators based on GPU microarchitecture GP100 Pascal.
NVIDIA says that the peak performance of a single node for Tesla P100 PCIe is 4.7 teraflops in double precision operations, and 9.3 teraflops in single precision operations, and, finally, 18.7 teraflops in operations with half precision. Also note that in the version of the accelerator for servers NVLink these figures are 5.3, 10.6 and 21 teraflops, respectively. According to NVIDIA, these figures correspond to the performance of more than 32 servers based on conventional CPU (which is, of course, is not specified).
Graphic processor for Tesla P100 PCIe has 3584 stream processors and works at a frequency of 1300 MHz (GPU Boost). The accelerator will be available in two versions, with 12 and 16 GB of video memory CoWoS HBM2 with a frequency of 1400 MHz and support for ECC. In the former case, memory bandwidth is 540 GB / s, and in the second - 720 GB / s. Level TDP for accelerators Tesla P100 PCIe 250 watt, and for heat dissipation radiator meets two slots high.
NVIDIA Tesla P100 Graphic Accelerator for PCIe-systems in the fourth quarter of 2016 will be on sale, as well as in the finished server makers such as Cray, Dell, Hewlett Packard Enterprise, IBM and SGI. Related Products :
|