NVIDIA DGX

NVIDIA DGX is a line of purpose-built AI supercomputers developed by NVIDIA, combining high-end GPUs with optimised interconnects, pre-installed AI software and support into turnkey systems for training and running large models. Introduced in 2016, the family has tracked each generation of NVIDIA hardware — Pascal, Volta, Ampere, Hopper and Blackwell — and includes rack-scale data-centre systems, the cluster-scale DGX SuperPOD architecture, and the desktop-sized DGX Spark.^[1][2]

History

The first DGX-1 was unveiled in April 2016 as what NVIDIA called the world's first AI supercomputer in a box: eight Pascal P100 GPUs delivering about 170 teraflops, with an early unit delivered to OpenAI later that year.^[1] Generations followed roughly in step with NVIDIA's GPU roadmap: the DGX-1 with Volta V100 in 2017, the 16-GPU DGX-2 in 2018, the DGX A100 in 2020 — the first to support Multi-Instance GPU partitioning — and the DGX H100 in 2022, built on the Hopper architecture with FP8 precision and a Transformer Engine aimed at large language models.^[1][4] The Blackwell era began with the DGX B200 in 2024 and the more capable DGX B300, alongside Grace Blackwell rack-scale systems such as GB200 NVL72 used at cluster scale.^[1] In March 2025, NVIDIA extended the line downward with the DGX Spark, a compact desktop system built on the GB10 Grace Blackwell superchip with 128 gigabytes of unified memory for local model work, and the workstation-class DGX Station. NVIDIA has signalled a move to the Rubin architecture from late 2026.^[1][3]

Key Concepts and Technology

DGX systems pair GPU clusters with high-bandwidth memory, NVLink interconnects between chips and high-speed InfiniBand networking between systems, all running the Ubuntu-based DGX OS with NVIDIA's AI software stack pre-installed. The design philosophy is integration: rather than assembling and tuning a generic server cluster, customers deploy systems NVIDIA says are ready for large training runs out of the box. Scaling is achieved through the DGX SuperPOD reference architecture, which links many systems into a single fabric, and the same capability can be rented through DGX Cloud, NVIDIA's cloud delivery model.^[1][2]

Applications and Impact

DGX systems are used by enterprises, national laboratories, universities and sovereign AI programmes to train and fine-tune foundation models, run large-scale simulations and serve inference at scale. Their releases are often treated as markers of the frontier of available compute, and the line set the template that much of the industry's AI data-centre design follows. The desktop Spark brought the format to individual developers; its early reception was turbulent, with developer John Carmack publicly reporting that his unit drew roughly 100 watts and delivered about half the quoted performance, drawing wide coverage before NVIDIA shipped updates that reviewers said improved the platform.^[1][3][7][8]

>See Also

• NVIDIA
• NVIDIA Blackwell
• GPU Cluster
• High Bandwidth Memory
• YTL AI Cloud