Nvidia Unveils First Volta-Based GPU

Print

Nvidia presents the Tesla V100-- the first GPU based on the company's Volta architecture, packing 3840 CUDA cores and 15 billion transistors on a 815mm slab of silicon.

Nvidia VoltaBuilt using a 12-nanometer manufacturing process, the Tesla V100 is aimed at high performance computing applications. As the successor to the current Pacal GPU flagship, the Tesla P100, it features a redesigned streaming architecture promising a 50% increase in efficiency compared to Pascal, enabling "major boosts in FP32 and FP64 performance in the same power envelope."

In addition the GPU carries 672 tensor cores (TCs)-- a new kind of core designed for machine learning operations. The company claims these boost performance by 4x over Pascal, and make the V100 superior to the Google dedicated tensor processing unit (TPU).

In terms of raw numbers, the V100 features 7.5 TFLOP of double precision floating-point (FP64) performance, 15 TFLOPs of single precision (FP32) performance and 120 Tensor TFLOPs of mixed-precision matrix-multiply-and-accumulate. Such power is paired with 16GB of 4096-bit HBM2 memory and a 2nd generation version of NVLink technology allowing transfer speeds of up to 300GB/s. Clock speed reaches 1455MHz, while TDP is rated at 300W.

The Tesla V100 will first ship in an Nvidia compute server-- a DGX-1 rack-mount number packing eight cards. The server should ship on Q3 2017, followed by 250W PCIe slot and half-height 150W versions of the card.

Go Nvidia Tesla V100