Japanese Arm-Powered Supercomputer Takes the TOP500 Crown

Twice a year, around the middle of the year (that would be now) and around the end of the year, the Top 500 List is published. This lists the 500 most powerful supercomputers in the world. In one sense, this is as much about bragging rights as anything else. After all, how "powerful" a supercomputer is depends very much on what you want to do with it. Weather forecasting, high-energy physics, playing chess, and training huge neural networks all require a lot of computation but stress different attributes of the machine. 

The list is used as a sort of definition of what is a supercomputer. After all, your smartphone would have been on the list if you could take it back in time a few decades. So the working definition is that it is a supercomputer if it's on the list in the top 500.

I've written about the supercomputer list a couple of times before. A post from December 2017, Supercomputers, covers the list in general and the #1 supercomputer at that time was the Chinese Sunway TaihuLight (with 10.6 million cores using a proprietary processor architecture).

Then, in August 2018, in a post Deep Learning and the Cloud I talked about Summit, the then-current #1 supercomputer, at the Oak Ridge National Laboratory. I said in that post:

Each node in Summit has two 22-core IBM Power9 CPUs and six NVIDIA Tesla V100 accelerators. But there are 4,608 of these nodes (that number turns out to be 4,096 + 512, so it's not quite as weird as it looks to computer scientists who only count in powers of 2). So I make that 202,752 Power9 cores, and 27,648 NVIDIA Volta GPUs. There's also a little memory...10 petabytes, along with 250 petabytes of storage. Peak performance is 200 petaflops. The DoE is planning an exaflops level machine for 2021.

As you can see, the bragging about being top of the charts is not limited to the institutions that built the computers, or even the architectures of the processors, whole countries get involved.

Fugaku

Well, there's a new country and a new architecture at #1. The highest performance supercomputer in the world is the Japanese Arm-based Fugaku. Apparently Fugaku is another name for Mount Fuji, but also presumably picked since it was made by Fujitsu. Maybe we should count Arm as a Japanese company since it's owned by Softbank. I still consider it English because all the early work was done in a barn outside Cambridge by people, many of whom I know.

As it says in the top500 press release announcing the new list:

The new top system, Fugaku, turned in a High Performance Linpack (HPL) result of 415.5 petaflops, besting the now second-place Summit system by a factor of 2.8X. Fugaku, is powered by Fujitsu’s 48-core A64FX SoC, becoming the first number one system on the list to be powered by Arm processors. In single or further reduced precision, which are often used in machine learning and AI applications, Fugaku’s peak performance is over 1,000 petaflops (1 exaflops). The new system is installed at RIKEN Center for Computational Science (R-CCS) in Kobe, Japan.

Summit is #2. Third is Sierra at Lawrence Livermore Laboratory (just over the hill from where I sit typing this in San Jose). Fourth is TaihuLight that I mentioned above.

Full details of all the top 10 are in the press release, and the whole list is on the top500 website.

Technology

As you might guess from the first two letters of its name, Fugaku was built by Fujitsu. It is the first time an Arm-based supercomputer has been at the top.

The press release gives a summary of the breakdown of technologies:

A total of 144 systems on the list are using accelerators or coprocessors, which is nearly the same as the 145 reported six months ago.  As has been the case in the past, the majority of the systems equipped with accelerator/coprocessors (135) are using NVIDIA GPUs.

The x86 continues to be the dominant processor architecture, being present in 481 of the 500 systems. Intel claims 469 of these, with AMD installed in 11 and Hygon in the remaining one.  Arm processors are present in just four TOP500 systems, three of which employ the new Fujitsu A64FX processor, with the remaining one powered by Marvell’s ThunderX2 processor.

The breakdown of system interconnect share is largely unchanged from six months ago. Ethernet is used in 263 systems, InfiniBand is used in 150, and the remainder employ custom or proprietary networks. Despite Ethernet’s dominance in sheer numbers, those systems account for 471 petaflops, while InfiniBand-based systems provide 803 petaflops. Due to their use in some of the list’s most powerful supercomputers, systems with custom and proprietary interconnects together represent 790 petaflops.

Fugaku Video

Here is a video titled Supercomputer "Fugaku" optimized for application performance and energy efficiency. There is a fair bit of information about the development and the processor. It is a standard ArmV8-A instruction set with SVE (scalable vector extension) and uses HBM2 memories.

Here's one slide from the video that is the most interesting part from a semiconductor point of view, but the full video (about half-an-hour) is worth a watch.

Sign up for Sunday Brunch, the weekly Breakfast Bytes email.