We’ve come to expect Nvidia CEO Jensen Huang’s annual SC keynote to contain stunning graphics and lively bravado (with plenty of examples) in support of GPU-accelerated computing. In recent years, AI has joined the show. SC19 didn’t disappoint. One of this year’s images was a gorgeous yellow Lamborghini rendered by a single GeForce RTX graphics card (1 GPU) – a task that would have required a cluster of GPUs a few years ago. (I wanted to jump in and drive away).

Let’s not forget computer graphics was there at the start and is still here in the heyday of GPUs reminded Huang, who has the rare skill to mix entertainment with key announcements and pull it off.

Not prone to understatement, Huang said, “HPC is literally everywhere today. It’s in supercomputing centers. It’s in the cloud. It’s at the edge. It is the most important and the most exciting developments that we’re seeing in computing today.” With that, he was off and running. It was part history, part future, but all Huang walking through Nvidia’s products and roadmap and greatest hits (so far).

Much of Huang’s keynote covered familiar ground and an hour and 54 minutes a lot to capture. Here’s a link to the video.

That said, amid the showmanship there was substantial news.

• Arm reference platform. Nvidia announced roll-out of a reference design for building GPU-accelerated Arm-based servers. The lack of a clear accelerator strategy has been a major stumbling block for Arm push to penetrate HPC and enterprise datacenters. Converged HPC/AI applications, for example, rely heavily on accelerators and heterogeneous compute architecture. Nvidia’s recent support is a major endorsement.
• Nvidia-Azure Supercomputer. Microsoft now offers a “supersized” NDv2 instance with up to 800 Nvidia V100 Tensor Core GPUs interconnected on a single Mellanox InfiniBand backend network. Nvidia likened it to being able to “rent an entire AI supercomputer on demand from their desk, and match the capabilities of large-scale, on-premises supercomputers that can take months to deploy.”
• Magnum IO. It’s a suite of software optimized to eliminate storage and input/output bottlenecks. Nvidia says Magnum IO delivers up to 20x faster data processing for multi-server, multi-GPU computing nodes when working with massive datasets to carry out complex financial analysis, climate modeling and other HPC workloads. At the heart of Magnum IO is GPUDirect, which provides a path for data to bypass CPUs and travel on “open highways” offered by GPUs, storage and networking devices.

The Arm news was perhaps the most significant. Arm seemed everywhere at SC this year. There was a packed birds-of-feather meeting today – Can Arm Take Leadership in HPC –  with Nvidia among the many presenters. Stir in any of several recent Arm advances – one notable is the new Cray-Fujitsu collaboration to develop a commercial supercomputer based on Fujitsu A64FX Arm-based processor going into the post-K “Fugaku” supercomputer – and Arm’s high-end aspirations are suddenly looking bright.

Nvidia has collaborated widely in its recent efforts to bolster (and become part of) the Arm community.

“To build the reference platform, NVIDIA is teaming with Arm and its ecosystem partners — including Ampere, Fujitsu and Marvell — to ensure NVIDIA GPUs can work seamlessly with Arm-based processors. The reference platform also benefits from strong collaboration with Cray, a Hewlett Packard Enterprise company, and HPE, two early providers of Arm-based servers,” reported Nvidia.

Additionally, “a wide range of HPC software companies have used NVIDIA CUDA-X libraries to build GPU-enabled management and monitoring tools that run on Arm-based servers,” according to Nvidia.

Back in the spring Nvidia announced it would bring CUDA to Arm. At SC19 the company made good on the promise saying it was now previewing its Arm-compatible software development kit, consisting of NVIDIA CUDA-X libraries and development tools for accelerated computing.

“There is a renaissance in high performance computing,” Huang said. “Breakthroughs in machine learning and AI are redefining scientific methods and enabling exciting opportunities for new architectures. 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.”

Arm chips, of course, are among the most-used in the world for many embedded and small devices applications. Advocates have always cited lower power as an advantage over x86. Gopal Hegde, VP/GM, Server Processor Business Unit, for Arm chip designer Marvel which collaborated with Nvidia on the Cuda work, estimates the power reduction advantage is 15-to-18 percent on comparable chips. Given the worry over power budgets for big systems that could become a significant factor processor selection decisions. Also, many of the prior ecosystem shortcomings (tools, apps, etc) have been or will soon be solved.

Several supercomputing centers are now standing up or testing systems with Arm chips and virtually all of use accelerators of one or another kind.

Oak Ridge National Laboratory science director Jack Wells provided the following assessment: “With the help of HPE, Marvell and NVIDIA, we’re excited to see how quickly our Oak Ridge National Laboratory-led effort could upgrade our Arm test-bed system, pull together performance testing, and get positive results. In just two weeks we were able to compile and correctly run approximately eight leadership-class applications, three important community libraries, and benchmark suites frequently used for evaluating the Arm-HPC ecosystem. Based on early results, the functionality of this Arm-hosted accelerated computing ecosystem appears to be already on par with the POWER and x86 environments.”

With Intel finally getting into the GPU game in earnest (officially announced at SC19), it is of course in Nvidia’s interest to expand its footprint.

The Azure news was interesting. Nvidia GPUs, of course, are in many clouds. Huang emphasized during his talk that HPC is moving quickly to the cloud and being asked to tackle ever larger job citing done as part a project to detect neutrinos (Ice Cube) in which 50K Nvidia GPUs were used. He also argues that shortened time-to-completion for jobs, not CPU or GPU instance rates, was the metric of value and would be less expensive.