“The world of computing has changed,” announced Nvidia founder and CEO Jensen Huang this week as he unveiled the new NVIDIA HGX-2 Cloud Server Platform at his company’s GPU Technology Conference in Taiwan.
The HGX-2 is the first platform fusing High Performance Computing (HPC) and AI within a unified architecture. Nvidia’s NVSwitch interconnect system enables 16 Nvidia Tesla V100 Tensor Core GPUs to perform as one giant GPU, delivering AI computing with a speed of 2 petaflops, or 2 quadrillion floating point operations per second.
The HGX-2 not only performs high-precision calculations using FP64 and FP32 for scientific computing and simulations, but also supports F16 and Int8 for AI training and inference. Nvidia says a single HGX-2 can replace 300 CPU-powered servers, bringing data centres significant savings in cost, space, and energy.
Following on the HGX-2 release, top server makers Lenovo, QCT, Supermicro and Wiwynn are expected to launch their own HGX-2 based systems later this year. Meanwhile, world-class original design manufacturers (ODMs) Hon Hai Precision Industry (known as Foxconn), Inventec, Quanta, and Wistron are also evaluating HGX-2 for their cloud data centres.
Author: Robert Tian| Editor: Tony Peng, Michael Sarazen