Server manufacturers such as Foxconn, Inventec, Quanta, and Wistron
Plan to build AI system for hyperscale data centers utilizing NVIDIA HGX

NVIDIA announced the NVIDIA HGX Partner Program, a partner program with global ODM companies, at Computex, a technology exhibition held in Taiwan. Through this new partner program, designed to respond more quickly to the demand for artificial intelligence cloud computing, NVIDIA plans to work closely with Foxconn, Inventec, Quanta, and Wistron.

First, NVIDIA plans to release the NVIDIA HGX reference architecture, NVIDIA GPU computing technology, and design guidelines to each ODM in advance. HGX is the same data center design used in Microsoft's 'Project Olympus,' Facebook's 'Big Basin' system, and the NVIDIA DGX-1 AI supercomputer.

As the overall demand for artificial intelligence computing resources increased rapidly last year, the performance and market adoption of NVIDIA's GPU computing platform also increased dramatically. Currently, all of the world's top 10 hyperscale companies are using NVIDIA GPU accelerators in their data centers.


By utilizing GPUs based on the new NVIDIA Volta architecture, which offers about three times better performance than existing products, ODM companies can respond to market demand with new products based on the latest NVIDIA technology.

NVIDIA has developed the HGX reference design to provide the high performance, efficiency, and massive scalability essential for hyperscale cloud environments. HGX allows for easy configuration changes to suit workload needs, enabling the easy combination of GPUs and CPUs according to the various methods required for high-performance computing, deep learning training, and inference.

The standard HGX design architecture includes eight NVIDIA Tesla GPU accelerators in an SXM2 form factor, connected to a cube mesh using NVIDIA NVLink high-speed interconnects and optimized PCIe topologies. Adopting a modular design, it can be deployed in any data center rack worldwide and utilize hyperscale CPU nodes when needed.

Both the NVIDIA Tesla P100 and V100 GPU accelerators are compatible with HGX. Therefore, when the V100 GPU is released in the second half of this year, all HGX-based products can be upgraded immediately.

NVIDIA GPU Cloud Platform can manage fully integrated deep learning framework containers optimized for NVIDIA GPU Cloud Platform, such as Caffe2, Cognitive Toolkit, MXNet, and TensorFlow.

Officials from each company expressed their expectations, stating, “An AI-based cloud environment will enable efficient work and provide groundbreaking solutions for computing requirements.”

Kushagra Vaid, General Manager and Principal Engineer for Microsoft’s Azure Hardware Infrastructure, said, “The HGX AI accelerator, developed as a component of Microsoft’s Project Olympus, achieves ultimate performance scalability through high-bandwidth interconnection options that can connect up to 32 GPUs.”