With Arm, Nvidia Sees More Choice, Openness And Efficiency

While Nvidia is declining to comment on a report that it is interested in buying Arm, the GPU juggernaut has been making increasingly larger investments in Arm-based processors in recent years, from using them in SoCs to creating new blueprints for Arm-based servers. ‘They offer a choice to the customers,’ an Nvidia director told CRN last fall.

When a report emerged Wednesday that Nvidia is interested in buying Arm from SoftBank Group, it was far from the first sign that the GPU juggernaut has expressed an interest in what the British chip designer can offer in the world of compute.

Nvidia and Arm have declined to comment on the reports, but Nvidia has been making increasingly larger investments in Arm-based processors in recent years, from using them in the company‘s own chips for embedded applications to creating new blueprints for Arm-based servers.

[Related: Nvidia CEO: Mellanox Buy Will Drive New Data Center Architecture]

Last year, Nvidia made a major move to support Arm-based processors with a new reference architecture and ecosystem support for GPU-accelerated, Arm-based servers, specifically for high-performance computing (HPC). The chipmaker also made its CUDA-X libraries, GPU-accelerated AI frameworks and software development tools available to Arm developers.

At the time, Nvidia CEO Jensen Huang hailed the announcement as significant for the HPC world, saying that “bringing Nvidia GPUs to Arm opens the floodgates for innovators to create systems for growing new applications from hyperscale cloud to exascale supercomputing and beyond.”

Among the supporters of Nvidia‘s new HPC alliance with Arm were Hewlett Packard Enterprise, Linux distributors Red Hat, SUSE and Canonical, and semiconductor companies Ampere Computing, Fujitsu and Marvell Technology Group, which license Arm‘s silicon IP to develop their own processors.

In an interview with CRN in the fall, Paresh Kharya, senior director of product management for accelerated computing at Nvidia, said Nvidia made a “big investment” to support Arm-based servers for three major reasons: more choice, an open platform and improved energy efficiency for customers.

Nvidia already supported x86 processors from Intel and AMD as well as IBM‘s Power processors, so it made sense that the chipmaker would extend its support to Arm-based processors as well, Kharya said, especially since customers were expressing interest.

“They offer a choice to the customers, and we are seeing customers throughout the world asking and are interested in Arm,” he said, pointing to interest in Japan and Europe for building supercomputers with Arm-based processors, like the new Fugaku system that debuted at the latest Top500 list among a few other HPC systems that use Arm.

With Arm‘s open platform, according to Kharya, customers can innovate on top of it to meet their specific compute requirements, which is what Nvidia itself has done with its Xavier embedded system-on-chips that come with Arm-based processors. Before Nvidia made its software stack available for Arm developers, the chipmaker was already using it for its Xavier SoCs, he added.

“For our own Xavier embedded SoCs, we chose Arm because it offered us the openness to innovate, to do the right ratios of how much sequential processing we need, how much AI processing we need, how much high-throughput processing we need,” he said. “And so we could create our own design with Arm-based architecture, and that‘s similar for the customers worldwide.”

Kharya also touted the energy efficiency benefits of Arm-based processors, pointing to Fugaku‘s ranking on Top500’s list of the 10 most energy-efficient systems in the world. Three other Arm-based systems are also on Top500’s current list of the 500 top energy- efficient supercomputers.

Arm‘s other major advantage is its pervasiveness, with Arm-based processors powering most of the world’s smartphones, including Apple’s iPhones, and many embedded devices.

“Arm is the most pervasive in instruction set in the world today. There are over a hundred billion devices with Arm, and they are known for their energy efficiency, and so that‘s one of the reasons why customers are exploring Arm as an instruction set,” Kharya said.

Khaya said he expects the adoption of Arm-based processors in the HPC space will likely lead to more widespread use in the data center market.

“HPC and supercomputing customers are generally the pioneers in innovating on the architectures and creating state-of-the-art breakthrough systems, so we expect them to be the first ones to start to experiment and explore this new GPU-accelerated architecture,” he said. “And we hope based on the success of supercomputing, this might further percolate into the enterprises.”

Eliot Eshelman, vice president of strategic accounts and HPC initiatives at Microway, a Plymouth, Mass.-based Nvidia partner, said while he‘s impressed by what Arm-based supercomputers like Fugaku are capable of, he doesn’t think it means the industry is ready for widespread adoption. Plus, there’s not enough incentive yet for most organizations to switch from x86-based systems to Arm-based ones, he added.

“We‘ve just seen that generally there’s enough natural resistance that there has to be a really good reason,” he said, “and one way I’ve heard it characterized is you kind of need a 10X incentive to change architecture. It needs to be 10 times cheaper or 10 times faster, or it may not be worth the level of effort, and we haven’t seen that yet from Arm seemingly.”

Microway has sent out a few proposals for Arm-based Marvell CPUs, but engagement has been “fairly limited” so far, according to Eshelman. But he thinks it‘s smart that Nvidia is making the investments in Arm now if the chipmaker knows Arm has long-term potential.

“If you look at how long it‘s taken them to get the [data center] industry to adopt their technology, depending on how you count, it’s probably been well over a decade. So if this is a multi-decade plan, and they’re willing to wait 10 years for most of the industry to adopt their new architecture, then maybe this is a great strategy,” Eshelman said.

As for whether Eshelman thinks it would make sense for Nvidia to buy Arm, he said it would allow the chipmaker to control the GPU, CPU and network in the data center and integrate them more tightly, expanding upon what Nvidia is already doing with its $8 billion acquisition of interconnect vendor Mellanox Technologies. But he acknowledged it could also make the compute ecosystem more messy since Arm has a different business model and licenses its silicon IP to many companies, including competitors like AMD.

“That is a huge wrinkle,” he said.