Skip to main content

Nvidia embraces ARM on its new Grace supercomputing processors

Image used with permission by copyright holder

After a lot of speculation, Nvidia finally announced its powerful supercomputer CPU called Nvidia Grace at its annual GPU Technology Conference. Named after famed American computer scientist and computer programming pioneer Grace Hopper, Nvidia’s first data-center processor is based on architecture from ARM. The company claimed in a press session ahead of GTC that Nvidia Grace is capable of delivering “10x the performance of today’s fastest servers on the most complex A.I. and high-performance computing workloads.”

This is Nvidia’s first partnership with ARM since announcing its $40 billion acquisition of the company in 2020.

Unlike Apple’s ARM-based M1 processor, Nvidia Grace won’t be headed to a computer near you. Instead, it is designed to be used on diverse applications like natural language processing, recommender systems, and A.I. supercomputing, Nvidia’s executives said. Nvidia Grace is adopted by both the Swiss National Supercomputing Centre (CSCS) and the United States Department of Energy Los Alamos National Laboratory to help power and drive scientific research.

Get your weekly teardown of the tech behind PC gaming
Check your inbox!

The company claimed that the volume of data and size of models used to train A.I. systems are growing exponentially, and the largest models are doubling every two-and-a-half months. To help reduce the bottleneck in communication between the processing and graphics core, Grace employs Nvidia’s NVLink interconnect technology, allowing large chunks of data to move between system memory, the CPU, and the GPU.

By tightly coupling the CPU and GPU inside Grace’s architecture, Nvidia stated that this processor will be seven times as fast on natural language training as Nvidia’s Selene supercomputer. Selene has a performance of 2.8-A.I. exaflops and is arguably one of the world’s fastest supercomputers for artificial intelligence applications today.

Grace combines energy-efficient ARM CPU cores with an innovative low-power memory subsystem.

And compared to traditional x86 server processors, like those made by rival Intel, Nvidia claimed that Grace is 10 times faster.

“A Grace-based system will be able to train a one trillion parameter NLP model 10x faster than today’s state-of-the-art NVIDIA DGXTM-based systems, which run on x86 CPUs,” the company added. This means that data that would have taken a month to analyze would now take just three days with Grace.

Powering this performance is ARM’s next-generation Neoverse cores alongside Nvidia’s fourth-generation NVLink interconnect, which has a bandwidth of 900 GB/s between the CPU and GPU. Grace will also use faster LPDDR5X memory, which is faster and more energy-efficient than DDR4 memory.

“It combines energy-efficient ARM CPU cores with an innovative low-power memory subsystem to deliver high performance with an energy-efficient design,” the company said.

Nvidia Grace will be used by Hewlett Package Enterprise in its Alps supercomputer, which will become available in 2023.

And Bluefield 3 for data centers too

Image used with permission by copyright holder

In addition to Nvidia Grace, the company also took the wraps off of its data center CPU called Bluefield 3. Referred to as the DPU, or data processing unit, Bluefield 3 is built for A.I. and accelerated computing that’s optimized for “multi-tenant, cloud-native environments, offering software-defined, hardware-accelerated networking, storage, security, and management services at data-center scale.”

The DPU delivers the equivalent of 300 cores to power data center services, which can free up the CPU to run other applications.

“Modern hyperscale clouds are driving a fundamental new architecture for data centers,” said Nvidia CEO Jensen Huang. “A new type of processor, designed to process data center infrastructure software, is needed to offload and accelerate the tremendous compute load of virtualization, networking, storage, security, and other cloud-native A.I. services. The time for BlueField DPU has come.”

Nvidia claims that Bluefield 3 delivers 10x the performance of the prior generation Bluefield 2, and the DPU is powered by 16x Arm X78 cores and has four times the acceleration for cryptography. The chip is the first 400 SeE/NDR DPU, Nvidia claimed, and it supports PCIe 5.0.

Additionally, Nvidia also built in network security to allow the DPU to detect and respond to cyber threats.

Bluefield 3 is expected to become available in early 2022.

Editors' Recommendations

Chuong Nguyen
Silicon Valley-based technology reporter and Giants baseball fan who splits his time between Northern California and Southern…
Nvidia may launch three new Super GPUs to fight back AMD
Three RTX 4080 cards sitting on a pink background.

Nvidia may be readying three new GPUs -- the RTX 4080 Super, RTX 4070 Ti Super, and the RTX 4070 Super. We haven't seen Super cards since Nvidia's RTX 20-series, but if this leak turns out to be true, they're coming back. Will they be worthwhile enough to rank high among the best graphics cards? It's hard to say, but they could help it compete against AMD's recent GPUs.

The information comes from hongxing2020, a frequent leaker in the GPU space. Nvidia already has a decent spread of GPUs between the RTX 4080, RTX 4070 Ti, and the RTX 4070. However, if a refresh to the Ada lineup is reportedly on the way, we might see some notable changes, but only if Nvidia decides to shake things up and use a different chip for at least two out of those three GPUs.

Read more
I tested AMD’s RX 7800 XT against Nvidia’s RTX 4070, and there’s a clear winner
AMD logo on the RX 7800 XT graphics card.

With the release of the Radeon RX 7800 XT and RX 7700 XT, AMD finally set foot in the mainstream gaming sector with its RDNA 3 lineup. Nvidia, its biggest rival, which also makes some of the best graphics cards, was the first to release competing cards with similar performance -- at least on paper. But are Nvidia's options better than AMD's in this generation?

The RX 7800 XT was, from the get-go, said to be the competitor to Nvidia's RTX 4070, but in reality, these GPUs differ both in price and performance. We've tested the RX 7800 XT and compared it to the RTX 4070, and we now know which of these two GPUs is the one to pick.
Pricing and availability

Read more
AMD’s two new GPUs significantly undercut Nvidia
AMD's Scott Herkalman presenting the RX 7800 XT graphics card.

AMD has just unveiled two new graphics cards aimed at gamers -- the Radeon RX 7800 XT and the RX 7700 XT. Set to rival some of the best graphics cards in this generation, the two GPUs are a way to bridge the gap in AMD's modest RDNA 3 lineup. AMD teases that both cards are suited for 1440p gaming, which could help them compete in the mainstream gaming market.

The prices are what stand out: the RX 7800 XT starts at $500, while the RX 7700 XT goes for $450. That significantly undercuts Nvidia's RTX 4070 at $600, though the RX 7700 XT is more expensive than the RTX 4060 Ti at $450.  Nvidia could really use some competition, so let's hope that AMD's new GPUs will be powerful enough to make a dent in Nvidia's market domination.

Read more