Nvidia details Grace CPU Superchip, an Arm-based server processor


Missed a Data Summit session? Watch on demand here.

Early last year, Nvidia unveiled the Grace processor, the company’s first data center ARM-based processor for AI and high-performance computing applications. But at the time, Nvidia declined to reveal a number of key details about Grace, including the number of transistors in the chip and where it might be used. That changed today at the company’s GTC event in March 2022, where Nvidia unveiled the Grace CPU Superchip, the first discrete chip in the Grace portfolio.

Nvidia’s Grace CPU chip packs 144 Arm cores into a single socket. Comprised of two CPU chips linked via Nvidia’s NVLink-C2C – a new chip-to-chip interconnect – it is designed to complement the Grace Hopper Superchip, which pairs a Grace processor with Nvidia’s recently announced Hopper architecture-based GPUs.

“A new kind of data center has emerged — AI factories that process and refine mountains of data to produce intelligence,” Nvidia co-founder and CEO Jensen Huang said in a press release. “The Grace CPU Superchip offers the best performance, memory bandwidth, and Nvidia software platforms in a single chip and will shine as the world’s AI infrastructure CPU.”

Arm-based architecture

The Grace CPU Superchip is based on Arm’s Neoverse family. First introduced in 2018, Neoverse integrates technologies from the Arm ecosystem that span from the edge to the cloud, especially those that focus on the data center.

The specific architecture is Arm’s Armv9, which debuted in May 2021. The Grace CPU Superchip – which Nvidia claims draws around 500 watts of power – features a memory subsystem comprised of LPDDR5x memory that delivers speeds of bandwidth of up to 1 terabyte per second.

Nvidia’s Grace CPU superchip.

Paired with Nvidia’s ConnectX-7 network cards, the Grace CPU Superchip can be configured in servers as CPU-only standalone systems or as GPU-accelerated servers with one, two, four, or eight Hopper-based GPUs. Nvidia says it is working with high performance computing, supercomputing and cloud customers on applications for the Grace CPU Superchip, which is expected to be available – along with the Grace Hopper Superchip – in the first half of 2023.

Emerging market

While the data center market is largely dominated by x86 processors made by Intel and AMD – Intel’s server market share is around 90% – the power efficiency of Arm chips has made them an increasingly popular alternative. more attractive to business customers. Amazon Web Services runs Arm-based chips like Graviton3 in its data centers, while supercomputers like Fujitsu’s Fugaku run Arm processors in combination with GPUs.

Nvidia previously announced that Grace would be installed for the first time in 2023 in supercomputers designed by Hewlett Packard Enterprise HPE for the Swiss National Center for Supercomputing and the US Department of Energy’s Los Alamos National Laboratory. And in 2025, Nvidia plans to introduce the next generation of the Grace architecture, named Grace Next, ahead of a new family of GPUs.

As analyst Roger Entner noted in a 2021 article for Fierce Electronics, several companies — including Marvell, Ampere, and Huawei — have been trying to make an impact with Arm chips in the server market. Internally, Amazon, Microsoft, Baidu, Alibaba, Facebook and others have designed custom Arm-based chips tailored to their own data centers and needs. But all these efforts were not crowned with success. Huawei spent around $3 billion funding a development platform and ecosystem around its Arm server chips, while Qualcomm was forced out of the space after trying to do the same.

“The Grace chip is designed with a general Arm core design available for any [Arm] licensed and not specifically customized to work with Nvidia GPUs,” Entner noted. “If Nvidia is allowed to buy Arm for $40 billion, the combined company would have plenty of reason to build specialized processors and an integrated architecture that will be significantly more efficient than standard ARM cores, giving the larger Nvidia an edge. more substantial on the server market.”

Of course, Nvidia ultimately declined to go ahead with its planned acquisition of Arm, citing regulatory pressures. Still, with its BlueField DPUs, which also leverage Arm-based chips for processing, Nvidia’s recent moves show it’s serious about making waves in a server processor market that could be worth $17.89 billion. by 2026.

“Advanced AI and data science are pushing today’s computing architecture beyond its limits – processing unthinkable amounts of data,” Huang said last year when unveiling the portfolio. Grace. “Using ARM licensed[IntellectualProperty[NvidiaaconçuGracecommeunprocesseurspécifiquementpourl’IAetleHPCàgrandeéchelleCoupléauGPUetauDPUGracenousoffrelatroisièmetechnologiefondamentalepourl’informatiqueetlacapacitéderé-architecturerlecentrededonnéespourfaireprogresserl’IANvidiaestmaintenantuneentrepriseàtroispuces[intellectualproperty[NvidiahasdesignedGraceasaCPUspecificallyforgiant-scaleAIandHPCCoupledwiththeGPUandDPUGracegivesusthethirdfoundationaltechnologyforcomputingandtheabilitytore-architectthedatacentertoadvanceAINvidiaisnowathree-chipcompany”[propriétéintellectuelle[NvidiaaconçuGracecommeunprocesseurspécifiquementpourl’IAetleHPCàgrandeéchelleCoupléauGPUetauDPUGracenousoffrelatroisièmetechnologiefondamentalepourl’informatiqueetlacapacitéderé-architecturerlecentrededonnéespourfaireprogresserl’IANvidiaestmaintenantuneentrepriseàtroispuces[intellectualproperty[NvidiahasdesignedGraceasaCPUspecificallyforgiant-scaleAIandHPCCoupledwiththeGPUandDPUGracegivesusthethirdfoundationaltechnologyforcomputingandtheabilitytore-architectthedatacentertoadvanceAINvidiaisnowathree-chipcompany”

VentureBeat’s mission is to be a digital public square for technical decision makers to learn about transformative enterprise technology and conduct transactions. Learn more


About Author

Comments are closed.