NVIDIA announces H200 at SC23

Background

NVIDIA sells its data-center GPUs to server makers in the form of HGX baseboards: multi-GPU boards that system builders mount in their own chassis, cooling and networking designs. The board that the Hopper family had been built around was the HGX H100, and by the time of the H200 launch a broad set of server makers and cloud operators had machines and software stacks designed around it.^[1] That existing base is the backdrop for the H200 announcement.

The announcement

NVIDIA announced the HGX H200 at SC23, the supercomputing conference held in Denver, on 13 November 2023.^[1] The company presented the board as a way of speeding up Hopper, the AI computing platform it was already selling. Ian Buck, NVIDIA's vice president of hyperscale and HPC, said the H200 made the industry's leading end-to-end AI supercomputing platform faster.^[1]

Partners, clouds and timing

Alongside the product, NVIDIA named the server partners that would be able to update existing systems with an H200: ASRock Rack, ASUS, Dell Technologies, Eviden, GIGABYTE, Hewlett Packard Enterprise, Ingrasys, Lenovo, QCT, Supermicro, Wistron and Wiwynn.^[1] It also said that Amazon Web Services, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure would be among the first cloud providers to deploy H200-based instances.^[1]

On compatibility, NVIDIA said HGX H200 boards in four-way and eight-way configurations are compatible with the hardware and software of HGX H100 systems.^[1] On availability, it said H200 systems were expected to ship from server makers and cloud providers starting in the second quarter of 2024.^[1]

Why it mattered

The emphasis of the launch was continuity rather than a break with what came before. A board that fits the hardware and software of existing HGX H100 systems gives operators a way to refresh a machine design without redesigning everything around it, and naming twelve server partners signalled that the upgrade path would be broadly available rather than limited to one vendor.^[1] The cloud list pointed at where many users would first meet the part, since rented instances reach customers long before they buy their own hardware.^[1] The stated shipping window, the second quarter of 2024, set the timetable against which those plans would be measured. For readers following the Hopper line, the H200 announcement is best read as a mid-generation step: same platform, same partners, faster systems, with general availability still months away.