| REVIEW | Not surprisingly, Nvidia at Supercomputing 19 of the General Assembly, held on published many news, which we will mention later. But it overlooked a perhaps one of the most interesting: a faster, lower-power next-generation graphics accelerator card. |
Multiple participants in multiple news sites and found this point, Nvidia confirmed to me that this is indeed a new card. Nvidia's "Volta" this generation Tesla GPU accelerator card had been eliminated in 2017, so the upgrade should have already expired.
V100S is currently only available PCI Express 3 interface, but is expected to eventually support Nvidia's SXM2 interface. Nvidia is SXM-socket design cards, PCIe cards with different, it does not need power supply is connected. SXM2 allows the GPU via the Nvidia NVLink (a high-bandwidth interconnection energy) with each other or communicate with the CPU, data transfer speed is ten times faster than PCIe.
With this card, Nvidia claims to have single-precision 16.4 TFLOPS, 8.2 TFLOPS double precision performance and Tensor Core up to 130 TFLOPS. This is only designed to improve V100 SXM2 than 4-5%, but increased from 16 to 17% than the PCIe V100 variant.
Memory capacity is maintained at 32 GB, but Nvidia added High Bandwidth Memory 2 (HBM2), in order to improve memory performance to 1,134 GB / s, and this ratio PCIe SXM2 have increased 26%.
Typically, the performance will also lead to increase in power, but here, the overall power PCIe card is 250 watts, the same as the previous generation PCIe card. Thus, under the same power consumption, the card can provide an additional 16-17% of computing performance, and a 26% increase memory bandwidth.
This switched: https://www.linuxprobe.com/nvidia-tesla-gpu.html