Japan Followed a Different Path to Build the World’s Fastest Supercomputer
[English Version, for Japanese Version, click here]
Custom Arm-based Processor, No GPUs, Very Power Efficient
A Japanese supercomputer called Fugaku has taken the #1 spot in the TOP500 list of the world’s fastest supercomputers. Its builders have taken a unique approach which reflects longer-term thinking weaved in with Japan’s national interests. Fugaku is one of only four systems out of the 500 fastest benchmarked systems on earth based on an Arm processor architecture. Arm-based architectures are moving upmarket from low-power use cases such as mobile phone processors, and as they do so they threaten incumbent processor companies Intel and AMD. Fugaku is also fairly unique for not using co-processors (GPUs) like those from Nvidia. The project to build Fugaku was funded by a Japanese government research institute called RIKEN, working in collaboration with Fujitsu Ltd. The last time Japan had a #1 ranked supercomputer was in 2011, and it was the RIKEN/Fujitsu predecessor to Fugaku, called “K”. In late 2015 rumors began to circulate that the next generation “post-K” machine would pivot to Arm-based architectures. In 2016, Softbank surprised the world with a deal to buy UK-based Arm Ltd. for $32B. Fugaku is a powerful demonstration of how well Arm-based processors can work when applied to supercomputing.
How Fast and Efficient is Fugaku?
Fugaku is a stunning 2.8 times faster than the prior US-based #1 machine by IBM called Summit at Oak Ridge National Laboratory. With the addition of #Fugaku into the 2020 list, Japan’s total performance share of the TOP500 list jumped from 7% to 24%.
Some systems have great performance but high power consumption.The Fugaku system has #1 performance and ranks #9 in power efficiency, which is an incredible accomplishment.That speaks to the viability of scaling such a system while maintaining reasonable operational costs with respect to power consumption, which is a bigger consideration in Japan than elsewhere. Fujitsu and the Japanese government are clearly “all-in” with Arm architectures.
Why are Supercomputers Important?
Let’s drill down into supercomputers. Supercomputers are used for innovative research, commercial endeavors, and for national security. Typical applications include climate modeling, physical simulations, artificial intelligence, machine learning, and drug discovery. Recently supercomputers around the world have been engaged in the race to find vaccines for COVID-19. Before the turn of the century #supercomputers were dominated by Cray, followed by a Control Data Corporation spin-out (now defunct), and then three Japanese companies Fujitsu, NEC, and Hitachi. Fast forward to today and the share of systems in the #TOP500 #supercomputing list is dominated by China, but keep in mind that Chinese vendors made a big, coordinated push to benchmark their systems in the TOP500 between 2015 and 2018. China’s numbers have flattened at 44% of the TOP500 systems as of mid-2018. Japan has been holding steady at around a 6% share of systems for 15 years, with a dip in 2008-2010 after the global financial crisis.
About two-thirds of Japan’s current 29 supercomputer deployments are Japanese vendor-led systems. Non-Japanese vendor systems installed in Japan are dominated by HPE and Cray. HPE acquired Cray in 2019, and are working with Oak Ridge National Laboratory in the US to deliver an AMD-processor based system in 2021 that will rival the RIKEN/Fujitsu supercomputer.
Fujitsu’s Global Opportunity?
Fujitsu has branded the Arm-based processor built for Fugaku as the A64FX, which is fine-tuned for the High Performance Computing (HPC) market, a subset of the broader server market. Will this new processor revive Fujitsu’s traditional server business globally? Regardless of what server segment it targets, technology alone isn’t enough. Fujitsu and other Japanese server vendors have already been pushed out of the mainstream server market due to higher overall system costs, less globalized software, weak sales and marketing, and high-touch support models that work better in Japan than in global markets. Based on TOP500 data, Japanese companies have deployed only 5 out of the 471 benchmarked supercomputers outside of Japan in the last several years, 3 in Germany (NEC), one in Taiwan (Fujitsu), and one in Australia (Fujitsu). To address more global HPC opportunities, at least as a technology supplier, new Fujitsu partner HPE/Cray offers the A64FX processors in their CS500 server and has made agreements to make it available globally through Cray including the HPE/Cray ecosystem of sales, software, and support. Fujitsu will continue to sell and service supercomputers in the Japanese market. Call it an obligation. In any case, it may be Softbank rather than Fujitsu which benefits the most financially from the publicity around Fugaku through their investment in Arm Ltd.
But It’s Good to be #1
Having the world’s fastest benchmarked supercomputer is good for Japan. It’s great timing to be winning accolades in highly efficient supercomputing architectures since technological ingenuity is an image Japan wants to portray around the Tokyo Olympics. More importantly, autonomous vehicles, robots, and all sorts of “edge” devices benefit from low power computing, and Arm’s simpler instruction set seems more capable of scaling up than Intel’s complex instruction set is of scaling down. Fugaku helps demonstrate how you can get incredible performance with lower power consumption and no GPUs. There will be thousands of Japanese software engineers, hardware engineers, and scientists who will elevate their skills as a result of the government funded RIKEN/Fujitsu project, and with the ever so gradual increase in Japan’s labor force mobility, such skills may make their way into commercial projects. However, for those commercial projects to see global traction, Japanese companies in rapidly evolving sectors need to adopt a more hybrid approach to managing global business, empowering highly skilled, industry knowledgeable leaders closer to the markets being served. It remains true that having the best technology or product still doesn’t ensure success in highly competitive global markets.