This site may earn affiliate commissions from the links on this folio. Terms of use.

HGX-1-Feature

Nvidia and Microsoft announced a new system form gene today at the Open up Compute Summit for the Open Compute Project. Unlike the venerable ATX standard, this one is designed for data centers and aimed at maximizing GPU performance as part of Microsoft's Projection Olympus initiative.

According to Nvidia, HGX-1 is designed "to meet the exploding demand for AI calculating in the cloud — in fields such as autonomous driving, personalized healthcare, superhuman vocalism recognition, information and video analytics, and molecular simulations."

Microsoft'south Project Olympus has been pulling in headlines, with hardware launches from Intel, AMD, Qualcomm, and now Nvidia besides. Each of these platforms is intended to accelerate a specific blazon of workload or scenario. Co-ordinate to Microsoft, Intel's work with Project Olympus fielded support for Skylake processors, with future versions expected to add support for FPGA accelerators or Intel Nervana solutions. AMD solutions are Naples-centric, as you might expect, and Qualcomm'southward focus on its own upcoming 48-core CPU.

HGX-1

An HGX-ane chassis

Nvidia'south Project Olympus platform will pack eight Pascal GPUs (GP100s) into a single chassis, all connected via NVLink, Microsoft said. Upward to 32 GPUs can be supported by linking four HGX-1 systems together (it isn't articulate which standard is used to link the systems themselves).

As Patrick Moorhead, of Moor Insights & Strategy points out, the ATX comparing represents how ambitious Nvidia is being with this push. The rollout of ATX in 1995 betwixt Microsoft and Intel gave the computer industry a single, unified course gene to design confronting. Information technology helped gear up the phase for an era where organization components could exist assumed to exist compatible with system chassis simply past conforming to the ATX standard. If the HGX-1 standard takes off similarly, hereafter HPC GPUs or CPUs would exist able to take advantage of the same kind of guarantees. The HGX-1 standard is designed to let CPUs and GPUs to connect in whatever ratio suits the workload, all via the NVLink interconnect.

Nvidia doesn't mention which CPUs this effort is compatible with, and it'll be interesting to come across if we see whatever AMD-Nvidia team-ups in this surface area in the futurity. AMD has its own server infrastructure and graphics division, but the RTG division inside AMD operates much more apart at present than information technology did in the by, and Nvidia has a vastly larger share of the HPC market than AMD does.

It might brand sense for AMD'due south server CPU team and Nvidia'southward HPC division to piece of work together to aggrandize the cloud calculating market place, even if the companies are rivals in other business segments. AMD announced its own AI products, dubbed Radeon Instinct, belatedly concluding year, merely has yet to announce any major hardware partners or system designs.