On the OCP World Summit, NVIDIA is providing a glimpse into the way forward for gigawatt AI factories.
NVIDIA will unveil specs of the NVIDIA Vera Rubin NVL144 MGX-generation open structure rack servers, which greater than 50 MGX companions are gearing up for together with ecosystem assist for NVIDIA Kyber, which connects 576 Rubin Extremely GPUs, constructed to assist growing inference calls for.
Some 20-plus business companions are showcasing new silicon, parts, energy methods and assist for the next-generation, 800-volt direct present (VDC) information facilities of the gigawatt period that can assist the NVIDIA Kyber rack structure.
Foxconn supplied particulars on its 40-megawatt Taiwan information middle, Kaohsiung-1, being constructed for 800 VDC. CoreWeave, Lambda, Nebius, Oracle Cloud Infrastructure and Collectively AI are amongst different business pioneers designing for 800-volt information facilities. As well as, Vertiv unveiled its space-, cost- and energy-efficient 800 VDC MGX reference structure, a whole energy and cooling infrastructure structure. HPE is asserting product assist for NVIDIA Kyber in addition to NVIDIA Spectrum-XGS Ethernet scale-across know-how, a part of the Spectrum-X Ethernet platform.
Transferring to 800 VDC infrastructure from conventional 415 or 480 VAC three-phase methods gives elevated scalability, improved power effectivity, diminished supplies utilization and better capability for efficiency in information facilities. The electrical automobile and photo voltaic industries have already adopted 800 VDC infrastructure for related advantages.
The Open Compute Venture, based by Meta, is an business consortium of a whole bunch of computing and networking suppliers and extra centered on redesigning {hardware} know-how to effectively assist the rising calls for on compute infrastructure.
Vera Rubin NVL144: Designed to Scale for AI Factories
The Vera Rubin NVL144 MGX compute tray gives an energy-efficient, 100% liquid-cooled, modular design. Its central printed circuit board midplane replaces conventional cable-based connections for sooner meeting and serviceability, with modular growth bays for NVIDIA ConnectX-9 800GB/s networking and NVIDIA Rubin CPX for massive-context inference.
The NVIDIA Vera Rubin NVL144 gives a serious leap in accelerated computing structure and AI efficiency. It’s constructed for superior reasoning engines and the calls for of AI brokers.
Its elementary design lives within the MGX rack structure and shall be supported by 50+ MGX system and element companions. NVIDIA plans to contribute the upgraded rack in addition to the compute tray improvements as an open normal for the OCP consortium.
Its requirements for compute trays and racks allow companions to combine and match in modular trend and scale sooner with the structure. The Vera Rubin NVL144 rack design options energy-efficient 45°C liquid cooling, a brand new liquid-cooled busbar for greater efficiency and 20x extra power storage to maintain energy regular.
The MGX upgrades to compute tray and rack structure increase AI manufacturing unit efficiency whereas simplifying meeting, enabling a fast ramp-up to gigawatt-scale AI infrastructure.
NVIDIA is a number one contributor to OCP requirements throughout a number of {hardware} generations, together with key parts of the NVIDIA GB200 NVL72 system electro-mechanical design. The identical MGX rack footprint helps GB300 NVL72 and can assist Vera Rubin NVL144, Vera Rubin NVL144 CPX and Vera Rubin CPX for greater efficiency and quick deployments.
If You Construct It, They Will Come: NVIDIA Kyber Rack Server Era
The OCP ecosystem can be getting ready for NVIDIA Kyber, that includes improvements in 800 VDC energy supply, liquid cooling and mechanical design.
These improvements will assist the transfer to rack server technology NVIDIA Kyber — the successor to NVIDIA Oberon — which can home a high-density platform of 576 NVIDIA Rubin Extremely GPUs by 2027.
The simplest method to counter the challenges of high-power distribution is to extend the voltage. Transitioning from a standard 415 or 480 VAC three-phase system to an 800 VDC structure gives numerous advantages.
The transition afoot allows rack server companions to maneuver from 54 VDC in-rack parts to 800 VDC for higher outcomes. An ecosystem of direct present infrastructure suppliers, energy system and cooling companions, and silicon makers — all aligned on open requirements for the MGX rack server reference structure — attended the occasion.
NVIDIA Kyber is engineered to spice up rack GPU density, scale up community measurement and maximize efficiency for large-scale AI infrastructure. By rotating compute blades vertically, like books on a shelf, Kyber allows as much as 18 compute blades per chassis, whereas purpose-built NVIDIA NVLink change blades are built-in on the again by way of a cable-free midplane for seamless scale-up networking.
Over 150% extra energy is transmitted via the identical copper with 800 VDC, enabling eliminating the necessity for 200-kg copper busbars to feed a single rack.
Kyber will develop into a foundational factor of hyperscale AI information facilities, enabling superior efficiency, effectivity and reliability for state-of-the-art generative AI workloads within the coming years. NVIDIA Kyber racks provide a approach for purchasers to cut back the quantity of copper they use by the tons, resulting in tens of millions of {dollars} in value financial savings.
NVIDIA NVLink Fusion Ecosystem Expands
Along with {hardware}, NVIDIA NVLink Fusion is gaining momentum, enabling corporations to seamlessly combine their semi-custom silicon into extremely optimized and broadly deployed information middle structure, decreasing complexity and accelerating time to market.
Intel and Samsung Foundry are becoming a member of the NVLink Fusion ecosystem that features {custom} silicon designers, CPU and IP companions, in order that AI factories can scale up rapidly to deal with demanding workloads for mannequin coaching and agentic AI inference.
- As a part of the just lately introduced NVIDIA and Intel collaboration, Intel will construct x86 CPUs that combine into NVIDIA infrastructure platforms utilizing NVLink Fusion.
- Samsung Foundry has partnered with NVIDIA to fulfill rising demand for {custom} CPUs and {custom} XPUs, providing design-to-manufacturing expertise for {custom} silicon.
It Takes an Open Ecosystem: Scaling the Subsequent Era of AI Factories
Greater than 20 NVIDIA companions are serving to ship rack servers with open requirements, enabling the longer term gigawatt AI factories.
- Silicon suppliers: Analog Gadgets, Inc. (ADI), AOS, EPC, Infineon, Innoscience, MPS, Navitas, onsemi, Energy Integrations, Renesas, Richtek, ROHM, STMicroelectronics and Texas Devices
- Energy system element suppliers: BizLink, Delta, Flex, GE Vernova, Lead Wealth, LITEON and Megmeet
- Information middle energy system suppliers: ABB, Eaton, GE Vernova, Heron Energy, Hitachi Power, Mitsubishi Electrical, Schneider Electrical, Siemens and Vertiv
Study extra about NVIDIA and the Open Compute Venture on the OCP World Summit, going down on the San Jose Conference Middle from Oct. 13-16.
