SC 2019 supercomputing conference, NVIDIA low - key release of the new top computing card. Tesla V100S

Tesla V100 was first released at GTC 2017 graphics conference in May 2017Based onNew Volta architecture, gv100 coreIt is manufactured by TSMC 12NM process and integrates 21 billion transistors, covering an area of 815 square mm. It has 5120 CUDA cores and 640 tensor cores, initially based on sxm2 (300gb / s nvlink bus),Soon added PCIe form(32GB / s PCIe bus).

More than two years later, the status of Tesla V100 is still unshakable, while the latest Tesla v100s goes further, with both core and memory speeds up, but power consumption remains unchanged.

Tesla v100s has only one form of PCIe expansion card,The double precision floating-point performance is 8.2tflops (trillions of floating-point calculations), the single precision floating-point performance is 16.4tflops, and the deep learning performance is 130tflops. Compared with the PCIe and sxm2 versions of Tesla V100, the performance is increased by up to 17% and 5%, respectively.

In the aspect of display and memory, hbm2 is still used,Capacity fixed at 32GBThe 16GB version is no longer available, the bit width is still 4096 bit, and the frequency is accelerated from 1.75ghz to 2.21ghz,Bandwidth also increased from 900gb / s to 1134gb / s.

But at the same time of a big increase in speed,The power consumption of Tesla v100s is still 250WObviously, both the manufacturing process and the core architecture are more mature.

In addition, from the on-site display,Tesla V100S