The spirit of Grace Hopper will stay on at NVIDIA GTC.
Accelerated techniques utilizing highly effective processors — named in honor of the pioneer of software program programming — might be on show on the international AI convention working March 18-21, able to take computing to the following stage.
System makers will present greater than 500 servers in a number of configurations throughout 18 racks, all packing NVIDIA GH200 Grace Hopper Superchips. They’ll type the biggest show at NVIDIA’s sales space within the San Jose Conference Heart, filling the MGX Pavilion.
MGX Speeds Time to Market
NVIDIA MGX is a blueprint for constructing accelerated servers with any mixture of GPUs, CPUs and knowledge processing items (DPUs) for a variety of AI, excessive efficiency computing and NVIDIA Omniverse purposes. It’s a modular reference structure to be used throughout a number of product generations and workloads.
GTC attendees can get an up-close have a look at MGX fashions tailor-made for enterprise, cloud and telco-edge makes use of, reminiscent of generative AI inference, recommenders and knowledge analytics.
The pavilion will showcase accelerated techniques packing single and twin GH200 Superchips in 1U and 2U chassis, linked by way of NVIDIA BlueField-3 DPUs and NVIDIA Quantum-2 400Gb/s InfiniBand networks over LinkX cables and transceivers.
The techniques help business requirements for 19- and 21-inch rack enclosures, and lots of present E1.S bays for nonvolatile storage.
Grace Hopper within the Highlight
Right here’s a sampler of MGX techniques now accessible:
- ASRock RACK’s MECAI, measuring 450 x 445 x 87mm, accelerates AI and 5G providers in constrained areas on the fringe of telco networks.
- ASUS’s MGX server, the ESC NM2N-E1, slides right into a rack that holds as much as 32 GH200 processors and helps air- and water-cooled nodes.
- Foxconn gives a collection of MGX techniques, together with a 4U mannequin that accommodates as much as eight NVIDIA H100 NVL PCIe Tensor Core GPUs.
- GIGABYTE’s XH23-VG0-MGX can accommodate loads of storage in its six 2.5-inch Gen5 NVMe hot-swappable bays and two M.2 slots.
- Inventec’s techniques can slot into 19- and 21-inch racks and use three totally different implementations of liquid cooling.
- Lenovo provides a spread of 1U, 2U and 4U MGX servers, together with fashions that help direct liquid cooling.
- Pegatron’s air-cooled AS201-1N0 server packs a BlueField-3 DPU for software-defined, hardware-accelerated networking.
- QCT can stack 16 of its QuantaGrid D74S-IU techniques, every with two GH200 Superchips, right into a single QCT QoolRack.
- Supermicro’s ARS-111GL-NHR with 9 hot-swappable followers is a part of a portfolio of air- and liquid-cooled GH200 and NVIDIA Grace CPU techniques.
- Wiwynn’s SV7200H, a 1U twin GH200 system, helps a BlueField-3 DPU and a liquid-cooling subsystem that may be remotely managed.
- Wistron’s MGX servers are 4U GPU techniques for AI inference and blended workloads, supporting as much as eight accelerators in a single system.
The brand new servers are along with three accelerated techniques utilizing MGX introduced at COMPUTEX final Might — Supermicro’s ARS-221GL-NR utilizing the Grace CPU and QCT’s QuantaGrid S74G-2U and S74GM-2U powered by the GH200.
Grace Hopper Packs Two in One
System builders are adopting the hybrid processor as a result of it packs a punch.
GH200 Superchips mix a high-performance, power-efficient Grace CPU with a muscular NVIDIA H100 GPU. They share lots of of gigabytes of reminiscence over a quick NVIDIA NVLink-C2C interconnect.
The result’s a processor and reminiscence advanced well-suited to tackle at the moment’s most demanding jobs, reminiscent of working giant language fashions. They’ve the reminiscence and pace wanted to hyperlink generative AI fashions to knowledge sources that may enhance their accuracy utilizing retrieval-augmented era, aka RAG.
Recommenders Run 4x Quicker
As well as, the GH200 Superchip delivers better effectivity and as much as 4x extra efficiency than utilizing the H100 GPU with conventional CPUs for duties like making suggestions for on-line buying or media streaming.
In its debut on the MLPerf business benchmarks final November, GH200 techniques ran all knowledge middle inference checks, extending the already main efficiency of H100 GPUs.
In all these methods, GH200 techniques are taking to new heights a computing revolution their namesake helped begin on the primary mainframe computer systems greater than seven a long time in the past.
Register for NVIDIA GTC, the convention for the period of AI, working March 18-21 on the San Jose Conference Heart and just about.
And get the 30,000-foot view from NVIDIA CEO and founder Jensen Huang in his GTC keynote.
