New Class of Accelerated, Environment friendly AI Methods Mark the Subsequent Period of Supercomputing

NVIDIA at this time unveiled at SC23 the following wave of applied sciences that may carry scientific and industrial analysis facilities worldwide to new ranges of efficiency and power effectivity.

“NVIDIA {hardware} and software program improvements are creating a brand new class of AI supercomputers,” mentioned Ian Buck, vice chairman of the corporate’s excessive efficiency computing and hyperscale information middle enterprise, in a particular tackle on the convention.

Among the methods will pack memory-enhanced NVIDIA Hopper accelerators, others a brand new NVIDIA Grace Hopper methods structure. All will use the expanded parallelism to run a full stack of accelerated software program for generative AI, HPC and hybrid quantum computing.

Buck described the brand new NVIDIA HGX H200 as “the world’s main AI computing platform.”

Image of H200 GPU system
NVIDIA H200 Tensor Core GPUs pack HBM3e reminiscence to run rising generative AI fashions.

It packs as much as 141GB of HBM3e, the primary AI accelerator to make use of the ultrafast expertise. Operating fashions like GPT-3, NVIDIA H200 Tensor Core GPUs present an 18x efficiency enhance over prior-generation accelerators.

Amongst different generative AI benchmarks, they zip by 12,000 tokens per second on a Llama2-13B massive language mannequin (LLM).

Buck additionally revealed a server platform that hyperlinks 4 NVIDIA GH200 Grace Hopper Superchips on an NVIDIA NVLink interconnect. The quad configuration places in a single compute node a whopping 288 Arm Neoverse cores and 16 petaflops of AI efficiency with as much as 2.3 terabytes of high-speed reminiscence.

Image of quad GH200 server node
Server nodes based mostly on the 4 GH200 Superchips will ship 16 petaflops of AI efficiency.

Demonstrating its effectivity, one GH200 Superchip utilizing the NVIDIA TensorRT-LLM open-source library is 100x sooner than a dual-socket x86 CPU system and practically 2x extra power environment friendly than an X86 + H100 GPU server.

“Accelerated computing is sustainable computing,” Buck mentioned. “By harnessing the ability of accelerated computing and generative AI, collectively we are able to drive innovation throughout industries whereas decreasing our affect on the setting.”

NVIDIA Powers 38 of 49 New TOP500 Methods

The newest TOP500 record of the world’s quickest supercomputers displays the shift towards accelerated, energy-efficient supercomputing.

Because of new methods powered by NVIDIA H100 Tensor Core GPUs, NVIDIA now delivers greater than 2.5 exaflops of HPC efficiency throughout these world-leading methods, up from 1.6 exaflops within the Might rankings. NVIDIA’s contribution on the highest 10 alone reaches practically an exaflop of HPC and 72 exaflops of AI efficiency.

The brand new record comprises the very best variety of methods ever utilizing NVIDIA applied sciences, 379 vs. 372 in Might, together with 38 of 49 new supercomputers on the record.

Microsoft Azure leads the newcomers with its Eagle system utilizing H100 GPUs in NDv5 cases to hit No. 3 with 561 petaflops. Mare Nostrum5 in Barcelona ranked No. 8, and NVIDIA Eos — which not too long ago set new AI coaching information on the MLPerf benchmarks — got here in at No. 9.

Displaying their power effectivity, NVIDIA GPUs energy 23 of the highest 30 methods on the Green500. They usually retained the No. 1 spot with the H100 GPU-based Henri system, which delivers 65.09 gigaflops per watt for the Flatiron Institute in New York.

Gen AI Explores COVID

Displaying what’s potential, the Argonne Nationwide Laboratory used NVIDIA BioNeMo, a generative AI platform for biomolecular LLMs, to develop GenSLMs, a mannequin that may generate gene sequences that intently resemble real-world variants of the coronavirus. Utilizing NVIDIA GPUs and information from 1.5 million COVID genome sequences, it might probably additionally quickly determine new virus variants.

The work received the Gordon Bell particular prize final yr and was educated on supercomputers, together with Argonne’s Polaris system, the U.S. Division of Vitality’s Perlmutter and NVIDIA’s Selene.

It’s “simply the tip of the iceberg — the longer term is brimming with potentialities, as generative AI continues to redefine the panorama of scientific exploration,” mentioned Kimberly Powell, vice chairman of healthcare at NVIDIA, within the particular tackle.

Saving Time, Cash and Vitality

Utilizing the most recent applied sciences, accelerated workloads can see an order-of-magnitude discount in system price and power used, Buck mentioned.

For instance, Siemens teamed with Mercedes to research aerodynamics and associated acoustics for its new electrical EQE automobiles. The simulations that took weeks on CPU clusters ran considerably sooner utilizing the most recent NVIDIA H100 GPUs. As well as, Hopper GPUs allow them to cut back prices by 3x and cut back power consumption by 4x (beneath).

Chart showing the performance and energy efficiency of H100 GPUs

Switching on 200 Exaflops Starting Subsequent 12 months

Scientific and industrial advances will come from each nook of the globe the place the most recent methods are being deployed.

“We already see a mixed 200 exaflops of AI on Grace Hopper supercomputers going to manufacturing 2024,” Buck mentioned.

They embody the huge JUPITER supercomputer at Germany’s Jülich middle. It could possibly ship 93 exaflops of efficiency for AI coaching and 1 exaflop for HPC purposes, whereas consuming solely 18.2 megawatts of energy.

Chart of deployed performance of supercomputers using NVIDIA GPUs through 2024
Analysis facilities are poised to change on a tsunami of GH200 efficiency.

Based mostly on Eviden’s BullSequana XH3000 liquid-cooled system, JUPITER will use the NVIDIA quad GH200 system structure and NVIDIA Quantum-2 InfiniBand networking for local weather and climate predictions, drug discovery, hybrid quantum computing and digital twins. JUPITER quad GH200 nodes will likely be configured with 864GB of high-speed reminiscence.

It’s considered one of a number of new supercomputers utilizing Grace Hopper that NVIDIA introduced at SC23.

The HPE Cray EX2500 system from Hewlett Packard Enterprise will use the quad GH200 to energy many AI supercomputers coming on-line subsequent yr.

For instance, HPE makes use of the quad GH200 to energy OFP-II, a complicated HPC system in Japan shared by the College of Tsukuba and the College of Tokyo, in addition to the DeltaAI system, which is able to triple computing capability for the U.S. Nationwide Heart for Supercomputing Purposes.

HPE can also be constructing the Venado system for the Los Alamos Nationwide Laboratory, the primary GH200 to be deployed within the U.S. As well as, HPE is constructing GH200 supercomputers within the Center East, Switzerland and the U.Okay.

Grace Hopper in Texas and Past

On the Texas Superior Computing Heart (TACC), Dell Applied sciences is constructing the Vista supercomputer with NVIDIA Grace Hopper and Grace CPU Superchips.

Greater than 100 world enterprises and organizations, together with NASA Ames Analysis Heart and Complete Energies, have already bought Grace Hopper early-access methods, Buck mentioned.

They be a part of beforehand introduced GH200 customers equivalent to SoftBank and the College of Bristol, in addition to the huge Leonardo system with 14,000 NVIDIA A100 GPUs that delivers 10 exaflops of AI efficiency for Italy’s Cineca consortium.

The View From Supercomputing Facilities

Leaders from supercomputing facilities around the globe shared their plans and work in progress with the most recent methods.

“We’ve been collaborating with MeteoSwiss ECMWP in addition to scientists from ETH EXCLAIM and NVIDIA’s Earth-2 venture to create an infrastructure that may push the envelope in all dimensions of huge information analytics and excessive scale computing,” mentioned Thomas Schultess, director of the Swiss Nationwide Supercomputing Centre of labor on the Alps supercomputer.

“There’s actually spectacular energy-efficiency positive factors throughout our stacks,” Dan Stanzione, govt director of TACC, mentioned of Vista.

It’s “actually the stepping stone to maneuver customers from the sorts of methods we’ve accomplished up to now to taking a look at this new Grace Arm CPU and Hopper GPU tightly coupled mixture and … we’re trying to scale out by in all probability an element of 10 or 15 from what we’re deploying with Vista after we deploy Horizon in a pair years,” he mentioned.

Accelerating the Quantum Journey

Researchers are additionally utilizing at this time’s accelerated methods to pioneer a path to tomorrow’s supercomputers.

In Germany, JUPITER “will revolutionize scientific analysis throughout local weather, supplies, drug discovery and quantum computing,” mentioned Kristel Michelson, who leads Julich’s analysis group on quantum info processing.

“JUPITER’s structure additionally permits for the seamless integration of quantum algorithms with parallel HPC algorithms, and that is necessary for efficient quantum HPC hybrid simulations,” she mentioned.

CUDA Quantum Drives Progress

The particular tackle additionally confirmed how NVIDIA CUDA Quantum — a platform for programming CPUs, GPUs and quantum computer systems often known as QPUs — is advancing analysis in quantum computing.

For instance, researchers at BASF, the world’s largest chemical firm, pioneered a brand new hybrid quantum-classical technique for simulating chemical substances that may protect people towards dangerous metals. They be a part of researchers at Brookhaven Nationwide Laboratory and HPE who’re individually pushing the frontiers of science with CUDA Quantum.

NVIDIA additionally introduced a collaboration with Classiq, a developer of quantum programming instruments, to create a life sciences analysis middle on the Tel Aviv Sourasky Medical Heart, Israel’s largest educating hospital.  The middle will use Classiq’s software program and CUDA Quantum operating on an NVIDIA DGX H100 system.

Individually, Quantum Machines will deploy the primary NVIDIA DGX Quantum, a system utilizing Grace Hopper Superchips, on the Israel Nationwide Quantum Heart that goals to drive advances throughout scientific fields. The DGX system will likely be linked to a superconducting QPU by Quantware and a photonic QPU from ORCA Computing, each powered by CUDA Quantum.

Logos of NVIDIA CUDA Quantum partners

“In simply two years, our NVIDIA quantum computing platform has amassed over 120 companions [above], a testomony to its open, revolutionary platform,” Buck mentioned.

General, the work throughout many fields of discovery reveals a brand new development that mixes accelerated computing at information middle scale with NVIDIA’s full-stack innovation.

“Accelerated computing is paving the trail for sustainable computing with developments that present not simply superb expertise however a extra sustainable and impactful future,” he concluded.

Watch NVIDIA’s SC23 particular tackle beneath.