Nvidia plans to enable its GPUs to support Arm CPUs used in data centers and supercomputers, advancing its ambitions in the high performance computing space. Nvidia said it would open its Cuda development tools, including its artificial intelligence and accelerated computing software, to the Arm ecosystem by the end of 2019.
Nvidia is the largest player in discrete GPUs, which have become standard for running a set of computations central to artificial intelligence. The company’s chips are plugged into data centers and supercomputers to serve as accelerators, running artificial intelligence and high performance computing chores faster than the CPU alone. Nvidia GPUs can currently connect to chips based on IBM’s Power and Intel’s x86 architectures.
By adding the ability to accelerate Arm processors, Nvidia will ensure that its GPUs can support all major CPU architectures used in servers. Arm, which SoftBank acquired for $32 billion in 2016, is increasing its investment in data center chip designs. Arm is aiming to loosen Intel’s stranglehold on the server chip business by giving customers including Marvell, Amazon and Ampere Computing the ability to build better Arm CPUs.
“Our customers are looking for high-performance, Arm-based processors to run their most demanding workloads," Renee James, Ampere's CEO, said in a statement. Nvidia GPUs working with Arm CPUs "will accelerate our work in building out the software ecosystem for Arm-based servers." Matt Murphy, Marvell's CEO, said that Nvidia bringing Cuda to Arm CPUs "offers breakthrough levels of energy efficiency and application performance."
Arm-based server chips have long been seen as a potential replacement for x86-based chips sold by Intel and Advanced Micro Devices, offering higher performance per dollar as well as lower power. That could help cut costs for cloud computing players such as Microsoft, Amazon and Google. Intel is the white whale of the data center segment, holding more than 95% market share in server chips, which can cost thousands of dollars each.
Arm processors are being used in many new supercomputers, including Cray’s Isambard and Japan’s Fugaku. Nvidia is also doubling down in the server chip business: the company is paying $6.9 billion to buy Mellanox Technologies, which sells Infiniband and Ethernet networking chips. Mellanox’s silicon can be dropped in data centers and supercomputers to stitch servers together with high bandwidth and reduced latency.
“As traditional compute scaling ends, power will limit all supercomputers,” Jensen Huang, Nvidia’s CEO, said. He said the combination of Nvidia's GPUs and Arm’s CPUs would drive the development of exascale supercomputers, which are projected to pump out a billion billion operations per second of performance. Simon Segars, Arm’s CEO, called Nvidia’s move to support Arm CPUs a "milestone" for the high performance computing sphere.