With Nvidia’s Arm-based Grace processor at its core, the corporate has launched a supercomputer designed to carry out AI processing powered by a CPU/GPU mixture.
The brand new system, formally launched on the Computex tech convention in Taipei the DGX GH200 supercomputer is powered by 256 Grace Hopper Superchips, know-how that may be a mixture of Nvidia’s Grace CPU, a 72-core Arm processor designed for high-performance computing and the Hopper GPU. The 2 are linked by Nvidia’s proprietary NVLink-C2C high-speed interconnect.
The DGX GH200 contains a large shared reminiscence house of greater than 144TB of HBM3 reminiscence linked by its NVLink-C2C interconnect know-how. The system is a simplified design, and its processors are seen by thier software program as as one big GPU with one big reminiscence pool, stated Ian Buck, vp and basic supervisor of Nvidia’s hyperscale and HPC enterprise unit.
He stated the system could be deployed and skilled with Nvidia’s assist in AI fashions that may require reminiscence past the bounds of what a single GPU helps. “We’d like a totally new system structure that may break by way of one terabyte of reminiscence with a purpose to practice these big fashions,” he stated.
Nvidia claims an exaFLOP of efficiency, however that’s from eight-bit FP8 processing. Now the vast majority of AI processing is being carried out utilizing 16-bit Bfloat16 directions, which might take twice as lengthy. A method of it’s you would have a supercomputer that ranks within the prime 10 of the TOP500 supercomputer listing and occupy a relatively modest house.
Through the use of NVLink as a substitute of normal PCI Specific interconnects, the bandwidth between GPU and CPU is seven occasions quicker and requires a fifth of the interconnect energy.
Google Cloud, Meta, and Microsoft are among the many first anticipated to realize entry to the DGX GH200 to discover its capabilities for generative AI workloads. Nvidia additionally intends to offer the DGX GH200 design as a blueprint to cloud service suppliers and different hyperscalers to allow them to additional customise it for his or her infrastructure. Nvidia DGX GH200 supercomputers are anticipated to be out there by the top of the 12 months.
Software program is included.
These supercomputers include Nvidia software program put in to offer a turnkey product that features Nvidia AI Enterprise, the first software program layer for its AI platform that includes frameworks, pretrained fashions, and growth instruments; and Base Command for enterprise-level cluster administration.
DGX GH200 is the primary supercomputer to pair Grace Hopper Superchips with Nvidia’s NVLink Change System, the interconnect that permits the GPUs within the system to work collectively as one. The earlier technology system maxed out at eight GPUs working in tandem.
To get to the full-sized system nonetheless requires vital data-center actual property. Every 15 rack-unit chassis holds eight compute nodes, and there are two chassis per rack (or pod in Nvidia parlance) together with NVswitch ethernet and IP connectivity. As much as eight of the pods could be linked for as much as 256 processors.
The system is air cooled even if Hopper GPUs draw 700 Watts of energy, which suggests appreciable warmth. Nvidia stated that it’s internally creating liquid-cooled methods and is speaking about it with prospects and companions, however for now the DGX GH200 is cooled by followers.
To this point, potental customers of the system aren’t prepared for liquid cooling, stated Charlie Boyle, vp of DGX methods at Nvidia. “There will probably be factors sooner or later the place we’ll have designs that must be liquid cooled, however we have been in a position to preserve this one on air,” he stated.
Nvidia introduced at Computex that the Grace Hopper Superchip is in full manufacturing. Techniques from OEM companions are anticipated to be delivered later this 12 months.
Copyright © 2023 IDG Communications, Inc.
