NVIDIA introduced the HGX-2 with a package of a powerful cloud server platform for AI and High-Performance Computing (HPC) workload for the improvement of its computer solutions.
The HGX-2 will help developers build more and larger complex of AI and HPC models while training them in a short period of time. The HGX-2 has a GPU baseboard which will be designated to the keybuilding block for the server platform. It includes baseboard hosts with eight V100 32GB Tensor Core GPUs and six NVSwitches. The HGX-2 has two Petaflops tensor core flops and a GPU memory size of 521 GB. It is always 6 NVLink and 16 GPU connected. The bisection bandwidth has 2,400B per seconds which converted to 48 NVLink.
The advantage of NVSwitch is that it is typically the highest performance server which equipped in cloud data centers the previously HGX-1 architechture.
The new cloud platform will be available later this year according to the official page of the company.