How much does the NVIDIA Tesla V100s performance increase?

NVIDIA's low-key launch at SC 2019 supercomputing conference reveals a comprehensive upgrade and high-end positioning for the Tesla V100s. This newly launched compute card is an enhanced version of the Tesla V100 that catches the eye with its superior performance and luxurious looks.

The Tesla V100S is derived from the Volta architecture's GV100 core and is housed in a 12nm TSMC process with 21 billion transistors integrated to create a massive 815 square millimeter chip. With 5,120 CUDA cores and 640 Tensor cores, it was initially offered in both SXM2 and PCIe form factors, but today only the PCIe expansion card is retained, providing a significant performance boost.The Tesla V100S delivers a double-precision floating-point performance boost of 8.2TFlops, a single-precision boost of 16.4TFlops, and a deep-learning performance increase of up to 130TFlops, an increase of 17% and 5% respectively compared to the base version.

In terms of graphics memory, the Tesla V100S adheres to the HBM2 standard, with a fixed capacity of 32GB and a bit width of 4,096-bits, but the frequency has been increased to 2.21GHz and the bandwidth has jumped to 1,134GB/s, which shows a significant advancement in technology.

Despite the significant performance increase, power consumption remains at 250W, showing NVIDIA's precision in energy efficiency control and maturity in architectural optimization while improving performance. The most striking cosmetic change is that the base color of the shoulders of the Tesla V100S has been changed to an earthy gold color, making it stand out from the crowd.

Overall, the low-profile release of Tesla V100s not only represents a technological leap, but also symbolizes NVIDIA's strength in high-end computing, demonstrating its dual improvement in performance and design.