Nvidia on Tuesday announced a new server platform, the HGX-2, designed to meet the needs of the growing number of applications that seek to leverage both high-performance computing (HPC) and artificial intelligence.
The platform, Nvidia says, is the first to offer high-precision computing capabilities to handle both HPC and AI workloads. It uses FP64 and FP32 for scientific computing and simulations, while enabling FP16 and Int8 for AI training and inference.
The HGX-2 server platform consists of a pair of baseboards. Each baseboard hosts eight V100 32 Gb Tensor Core GPUs. The 16 GPUs are fully connected through 12 NVSwitches to collectively deliver two petaflops of AI performance.
With the HGX-2 serving as a “building block,” Nvidia works with server manufacturers to build full server platforms that can be customized to meet the needs of different data centers. For example, manufacturers can put networking cables in the back of the server in the traditional enterprise style or in the front of the server for easier servicing. The first system built using HGX-2 is the DGX-2, Nvidia’s recently announced flagship server.
A single HGX-2, Nvidia says, can replace up to 300 CPU-only servers on deep learning training. It’s achieved record AI training speeds of 15,500 images per second on the ResNet-50 training benchmark.
The HGX-2 comes a year after the launch of the original HGX-1, which was adopted by various server manufacturers and companies operating large data centers like Amazon Web Services and Facebook.
Four server makers are already committed to bring HGX-2 systems to market this year: Lenovo, QCT, Supermicro and Wiwynn. Additionally, four original design manufacturers (Foxconn, Inventec, Quanta and Wistron) are designing HGX-2-based systems for large cloud data centers.
Nvidia has been building a family of GPU-accelerated server platforms, each with different combinations of GPUs, CPUs and interconnects to accommodate different AI and HPC workloads. For instance, the HGX-T2 is built for deep learning training for large AI models, the HGX-I2 is optimized for inference and the SCX is optimized for supercomputing applications.