While the Top500 list ranks the most powerful benchmarked supercomputers in the world, the Green500 reorders the list to rank the most energy-efficient supercomputers in terms of FLOPS per watt, seeking to decouple the idea of performance from sheer speed and shed light on efforts to drive efficiency in machines that tend to consume vast amounts of energy. Now, the Green500 has a new leader: the MN-3 system created by Japanese AI startup Preferred Networks.
MN-3 topped the Green500 list, delivering a whopping 21.1 gigaflops per watt. The system delivers 1.62 Linpack petaflops overall, earning it 394th place in the brand new Top500 list. MN-3, which uses Intel Xeon CPUs in combination with Preferred Networks’ in-house MN-Core dedicated accelerator, is the successor to Preferred Networks’ MN-2 system, which became operational in 2019 and uses Nvidia V100 GPUs. Preferred Networks doesn’t plan on selling its Green500-topping technology, however; MN-3 and its proprietary hardware are only available for the company’s own R&D efforts.
The 21.1 gigaflops per watt achieved by the MN-3 is simultaneously impressive and foreboding. Fugaku, the Fujitsu-Arm based machine which dominated the new Top500 list with a staggering 415.5 Linpack petaflops and placed 9th on the corresponding Green500 list, delivers two thirds of MN-3’s efficiency – 14.7 gigaflops per watt. Extrapolating out to an exascale system, even MN-3 would require 47 megawatts to produce one exaflops, while Fugaku would require 68 megawatts to do the same. For context, dozens of wind turbines would be needed to produce 47 megawatts of power, and the same amount of power could provide electricity for tens of thousands of average homes.
MN-3’s list-topping efficiency rating falls around 15 percent short of the 40-megawatt (max) envelope targeted by planned exascale systems like Aurora, Frontier and El Capitan, and does not approach the 20-megawatt goal for exascale systems set by the U.S. Defense Advanced Research Project Agency (DARPA) in the early 2010s (those goals, respectively, would require – at minimum – 25 and 50 gigaflops per watt). With Aurora (planned as the first U.S. exascale system) currently scheduled to be delivered in just a year and a half within a 40-megawatt footprint, system developers need to close that 15 percent gap – and do it on a system many times the size of MN-3.
Luckily, MN-3 is in powerful company. Nvidia’s newest AI supercomputer, Selene, delivered 20.52 gigaflops per watt, placing it just behind MN-3 in the Green500 list – but it also delivered 27.5 Linpack petaflops, placing it 7th on the Top500 list. Overall, in fact, half of the Top500’s top ten also appear in the Green500’s top ten.
“You see a lot of top ten systems from the top ten of the Top500 in the top ten of the Green500,” said Top500 list co-author Eric Strohmaier in the Top500 session at ISC 2020. “That’s a very nice result, too: it’s not only small systems which show up here, but very big ones, too. That’s an observation we have made for many years – that the HPC community likes to innovate from the top, so very large systems often are the first ones to receive new technology. And that continues to be the case.”
The Preferred Networks win also adds to Japan’s winning streak. The country is now home to both the Green500 leader (MN-3) and the Top500 leader (Fugaku), with Japan-based systems claiming three additional spots in the Green500’s top ten.
Overall power efficiency is also on the rise. “Looking at the history of the Green500, you see the different architectures and how they have improved their efficiency over time,” Strohmaier said. “[MN-3 being ahead of Arm-based systems like Fugaku is] a nice increase, but the more important increase is actually the blue dots at the bottom of this slide, which is showing you the increase in average power efficiency in the list. And that’s a continuous march upwards, even if not quite as steep as the peak efficiency.”
Header image: The MN-3 system. Image courtesy of Preferred Networks.