NVIDIA and main server producers have unveiled a collection of methods powered by the NVIDIA Blackwell structure, that includes Grace CPUs and superior networking infrastructure. This transfer is about to allow enterprises to construct AI factories and knowledge facilities, driving the subsequent wave of generative AI breakthroughs.
Throughout his keynote at COMPUTEX 2024, NVIDIA founder and CEO Jensen Huang has introduced that prime server producers comparable to ASRock Rack, ASUS, GIGABYTE, Ingrasys, Inventec, Pegatron, QCT, Supermicro, Wistron, and Wiwynn will ship a variety of AI methods. These methods, using NVIDIA GPUs and networking, will cater to cloud, on-premises, embedded, and edge AI functions.
Mr. Huang declared, “The following industrial revolution has begun. Corporations and nations are partnering with NVIDIA to shift the trillion-dollar conventional knowledge facilities to accelerated computing and construct a brand new kind of information middle – AI factories – to provide a brand new commodity: synthetic intelligence.” He emphasised that your entire business, from server and networking producers to software program builders, is making ready for Blackwell to speed up AI-powered innovation throughout all fields.
These new choices would cater to a wide range of functions, that includes configurations from single to multi-GPU setups, x86- to Grace-based processors, and air- to liquid-cooling applied sciences. Moreover, to expedite the event of methods in varied sizes and configurations, the NVIDIA MGX modular reference design platform now helps Blackwell merchandise. This contains the brand new NVIDIA GB200 NVL2 platform, designed to ship unmatched efficiency for mainstream giant language mannequin inference, retrieval-augmented technology, and knowledge processing.
Accelerated Computing Wants of Knowledge Facilities
The GB200 NVL2 is tailor-made for rising market alternatives, comparable to knowledge analytics, the place corporations make investments tens of billions of {dollars} yearly. The platform leverages high-bandwidth reminiscence efficiency by NVLink-C2C interconnects and devoted decompression engines within the Blackwell structure, accelerating knowledge processing by as much as 18 instances with 8 instances higher power effectivity in comparison with x86 CPUs.
To handle the various accelerated computing wants of world knowledge facilities, NVIDIA MGX offers a reference structure that enables server producers to construct over 100 system design configurations “shortly and cost-effectively.” Producers can begin with a primary system structure for his or her server chassis after which choose their GPU, DPU, and CPU to fulfill totally different workload necessities. Presently, greater than 90 methods from over 25 companions are both launched or in improvement, a major improve from the 14 methods from six companions the earlier yr. The MGX structure helps scale back improvement prices by as much as three-quarters and shortens improvement time by two-thirds, to simply six months.
Intel and AMD
Each AMD and Intel are supporting the MGX structure, with plans to introduce their very own CPU host processor module designs. This contains AMD’s next-generation Turin platform and Intel’s Xeon 6 processor with P-cores (previously codenamed Granite Rapids). These reference designs permit any server system builder to avoid wasting improvement time whereas making certain consistency in design and efficiency.
NVIDIA’s newest platform, the GB200 NVL2, leverages MGX and Blackwell, providing a scale-out, single-node design that allows varied system configurations and networking choices. This ensures seamless integration of accelerated computing into present knowledge middle infrastructure. The GB200 NVL2 is a part of the Blackwell product lineup, which additionally contains Blackwell Tensor Core GPUs, GB200 Grace Blackwell Superchips, and the GB200 NVL72.
NVIDIA’s intensive associate ecosystem contains TSMC, the world’s main semiconductor producer and an NVIDIA foundry associate, in addition to world electronics makers that present key elements to create AI factories. These embrace improvements in server racks, energy supply, cooling options, and extra from corporations comparable to Amphenol, Asia Very important Elements (AVC), Cooler Grasp, Colder Merchandise Firm (CPC), Danfoss, Delta Electronics, and LITEON.
This collaborative effort would permit for the speedy improvement and deployment of recent knowledge middle infrastructure to fulfill the wants of world enterprises. The infrastructure is additional accelerated by Blackwell know-how, NVIDIA Quantum-2 or Quantum-X800 InfiniBand networking, NVIDIA Spectrum-X Ethernet networking, and NVIDIA BlueField-3 DPUs in servers from main methods makers Dell Applied sciences, Hewlett Packard Enterprise, and Lenovo. Enterprises may entry the NVIDIA AI Enterprise software program platform, which incorporates NVIDIA NIM inference microservices, to create and run production-grade generative AI functions.
Blackwell Expertise in Taiwan
Mr. Huang additionally highlighted that Taiwan’s main corporations are quickly adopting Blackwell know-how to reinforce their AI capabilities. Chang Gung Memorial Hospital, Taiwan’s premier medical middle, plans to make use of the NVIDIA Blackwell computing platform to advance biomedical analysis and speed up imaging and language functions, enhancing medical workflows and affected person care.
Foxconn, one of many world’s largest electronics producers, is about to make use of NVIDIA Grace Blackwell to develop sensible resolution platforms for AI-powered electrical automobiles and robotics. In addition they intention to broaden language-based generative AI providers to offer extra personalised buyer experiences.
NVIDIA’s newest developments and partnerships would signify a major leap within the AI-driven transformation of information facilities, poised to revolutionize varied industries worldwide.
“ASUS is working with NVIDIA to take enterprise AI to new heights with our highly effective server lineup, which we’ll be showcasing at COMPUTEX,” stated Jonney Shih, Chairman at ASUS. “Utilizing NVIDIA’s MGX and Blackwell platforms, we’re capable of craft tailor-made knowledge middle options constructed to deal with buyer workloads throughout coaching, inference, knowledge analytics and HPC.”
“Our building-block structure and rack-scale, liquid-cooling options, mixed with our in-house engineering and world manufacturing capability of 5,000 racks monthly, allow us to shortly ship a variety of game-changing NVIDIA AI platform-based merchandise to AI factories worldwide,” stated Charles Liang, president and CEO at Supermicro. “Our liquid-cooled or air-cooled high-performance methods with rack-scale design, optimized for all merchandise primarily based on the NVIDIA Blackwell structure, will give clients an unimaginable alternative of platforms to fulfill their wants for next-level computing, in addition to a significant leap into the way forward for AI.”
