Nvidia GPUs and Arm processors marry in new server reference design

Nvidia GPUs and Arm-based processors will work together in a new reference design of hardware and software blocks that companies can build into high performance servers for cloud, edge, AI, storage and supercomputing.

Nvidia CEO Jensen Huang announced the concept at the Supercomputing 19 conference in Denver on Monday along with several technology partners including Arm, Ampere, Fujitsu and Marvell. Cray and HPE, two early providers of Arm-based servers, also collaborated.

“Bringing Nvidia GPUs to Arm opens the floodgates for innovators to create systems for growing applications from hyperscale-cloud to exascale supercomputing and beyond,” Huang said.

Nvidia announced in June that its CUDA-X software would run on Arm designs, and now Nvidia is making available its Arm-compatible software development kit of CUDA-X libraries and development tools for accelerated computing.

“Nvidia GPUs have practically become standard accelerators for AI and High Performance Computing, so it’s a natural move for Nvidia to embrace leading Arm providers,” said Karl Fruend, an analyst at Moor Insights & Strategy. The move to Arm is also noteworthy for Arm’s recent slow growth in HPC “after years of disappointing performance and adoption,” he added.

Arm partners are able to build servers with solid performance and features such as increased memory and I/O bandwidth, he added.

In one example, Marvell announced on Monday that Nvidia GPUs support its ThunderX family of Arm servers. The ThunderX2 is Marvell’s latest 64-bit Armv8-A based server processor, which is combined with Nvidia GPUs.

Elsewhere at the conference, Intel announced an exascale GPU code-named Ponte Vecchio that will be used in the Aurora supercomputer at Argonne National Lab.

Also at the conference, Nvidia introduced Magnum IO software to help data scientists and researchers process large amounts of data in minutes instead of hours. It delivers up to 20 times faster data processing for multi-server, multi-GPU nodes, based on benchmarks. Magnum IO was developed with DataDirect Networks, Excelero, IBM, Mellanox and WekaIO.

In addition, Nvidia announced a scalable supercomputer for Microsoft Azure cloud that is GPU accelerated. Azure’s new NDv2 cloud instance will run Nvidia V100 Tensor Core GPUs, interconnected on a single Mellanox InifiBand network. Customers will be able to run an entire AI supercomputer on demand from a desktop.

Microsoft’s pricing is $22 an hour for one instance of eight Nvidia V100 GPUs on the NDv2 cloud and can be scaled up.

RELATED: Nvidia launches new AI embedded module, runs the table on MLPerf