HGX

HGX is a family of GPU accelerator platforms developed by NVIDIA for data centers. It refers to open, modular baseboards and associated reference designs that are intended to serve as building blocks for AI, HPC, and high-density server deployments. Variants such as HGX-1 and HGX-2 have been used by NVIDIA and its ecosystem to enable scalable multi-GPU configurations within a single chassis.

Architecture and features of HGX designs are centered on providing a high-density, high-bandwidth foundation for GPU acceleration. The baseboards are designed to host multiple NVIDIA GPUs, along with CPUs, memory, storage, and networking components, to form a complete server solution. Interconnect options across GPUs in HGX designs typically emphasize rapid data transfer, with mechanisms such as NVLink and, in certain generations, NVSwitch to enable efficient multi-GPU communication. The platforms are engineered for effective cooling and power delivery to support sustained workloads in data centers.

History and usage: NVIDIA introduced HGX designs to standardize the construction of AI and HPC servers. The HGX reference designs have formed the internal platform for several NVIDIA-based systems, including DGX appliances, and have been adopted by various OEMs and cloud providers to accelerate deployment of AI training and inference workloads.

Impact: The HGX family helped standardize high-density GPU server platforms, enabling faster time-to-market for AI servers and enabling scalable, enterprise-grade accelerators in data centers. It reflects NVIDIA’s approach to modular, interoperable hardware that supports evolving GPU generations and workloads.