2020 was a watershed year for Arm in High-performance Computing (HPC). Fugaku, the system developed by RIKEN on the Arm-based A64FX processor from Fujitsu, captured the number one spot on the Top500 list in June. It delivered 415.5 petaflops or 2.8 times more sustained performance over the second most powerful system on the list. Fugaku also set records for HPCG, Graph500, and the AI benchmark, sweeping the performance criteria build captured the top spot on the Green 500 list. Fugaku is ahead of schedule and already being deployed to help the effort against Covid-19. Early uses include simulating how the virus is exhaled and circulates in the air.
Sandia National Labs and Cray, meanwhile, have announced plans to deploy systems based around Fujitsu’s A64FX processor. The new system is Sandia’s second Arm-based supercomputer. Astra, their first, is built around Marvell’s ThunderX2 processor and became the first Arm system on the top 500 when it debuted in 2018. One of the key findings in this effort has been that the software environment on Arm is top notch and competes with the best of the traditional offerings.
Speaking of software, our partners at Nvidia announced full support for CUDA 11 for Arm, creating a pathway for bringing together its performance-leading GPUs with our CPUs. The European processor initiative and SiPearl, meanwhile, announced it selected Neoverse V1 (formerly code-named Zeus), a next-generation Neoverse processor, for its upcoming exascale computer.
But, to paraphrase, these accomplishments are not the end, or the beginning of the end. It is more like the end of the beginning. High-performance computing is substantially reshaped in the coming years by a few powerful trends. By virtue of its technology and approach, Arm plays a unique role in overcoming the looming challenges.
Before we get to those trends, I would like to highlight an event being put on by the Arm HPC User’s Group (AHUG). We are bringing together Arm HPC solution vendors, on-site practitioners, and researchers to share their plans, experiences, and results. Check out AHUG’s YouTube channel starting on Wednesday November 4 to see a series of virtual talks. And please register for our live roundtable and Q&A with the vendors on November 9. And now back to the trends:
Thinking inside the socket. With the slowing of Moore’s law, advances in performance increasingly come through innovations in design at the microarchitectural level inside the socket.
This is one of the most influential trends and will result in greater emphasis on domain-specific logic implemented inside the chip. Arm and partners invented SVE further out we see progress in chiplet architectures for building highly complex processors containing thousands of cores. We see spatial computing for further extending the benefits of parallelism and computational storage concepts for reducing the energy penalty that comes with moving data. The next ten years will extraordinary.
Different systems for different challenges. One of the primary stories of the last ten years in supercomputing has been the rise of accelerators, specifically GPUs, and the benefits of bulk-offload of heavy computational tasks to SIMD/Vector engines. Arguably started with LANL’s RoadRunner system, the trend has blossomed into GPU-accelerated clusters and an associated effort in the software infrastructure to support workloads with off-load in mind.
The next logical step is on-socket logic specialized for particular markets and use cases. Imagine a system optimized for climate simulations and another for oil & gas. While they would share key micro-architectural uses, they enable a common logic moving into the core. As implementation costs fall, the potential for such advances becomes closer to reality. The handheld and mobile market has been going down this path for several generations. With the flexibility of Arm’s business model, customers can add logic next to the core. The strength of the software stack can be leveraged to provide a world-class environment to run the workloads and facilitate low-latency hand offs to logic when applicable.
HPC in the cloud. R&D budgets at universities and national labs are under incredible pressure. At the same time, hyperscale cloud providers like Amazon, Google, and Microsoft manage some of the most sophisticated computing infrastructures in the world. And they are at the forefront of areas of research like quantum computing. HPC in the cloud becomes a way to meet the demands of the world within the economic circumstances of our time.
HPC in the cloud will also greatly expand the market. Think of a bike manufacturer, or even a collegiate bicycling team, taking advantage of cloud HPC to run aerodynamic simulations. Or a shoe manufacturer using HPC to determine the optimal mix of materials for manufacturing.
Finally, on the topic of cloud and HPC, a convergence of infrastructure methodologies is taking place. The old-school methods of dedicated HPC systems are being questioned as the successes of hyperscale centers grows. Watch for efforts toward modernizing traditional HPC interfaces as the best of both environments merge to gain advantages for users. Industry analysts are claiming the ease of use and availability make this a market expansion for HPC, not merely moving workload out of the traditional HPC data centers.
And since we cannot meet in person to discuss these ideas, please reach out and tell me where you think the future of supercomputing might be going.
Register for our HPC Roundtable