That is, the first single system to have more than 1018 transistors, more than a quintillion, or 1,000,000,000,000,000,000 . It’s the Fugaku supercomputer in Japan:
It’s the world’s fastest machine as of November 2020, as defined by the benchmarks on the Top 500 list of supers, and is the first to break an exa-flop, doing more than 1018 floating point operations per second. It’s named after Mt Fuji, the most beautiful mountain in the world. It was built by Fujitsu using their own custom processor chips, and is operated by the RIKEN Center for Computational Science in Kobe, near Osaka. Here’s how its transistors are allocated:
|Processors||1.5 x 1015||160K chips with 48 cores and 8.8 billion transistors each|
|DRAM||44 x 1015||32 GB on each chip above|
|Local flash storage||132 x 1015||1.5 TB on each set of 16 chips|
|System flash storage||1350 x 1015||150 PB for the whole machine, and assuming single-bit-per-transistor NAND flash|
|Total||1.5 x 1018|
The whole machine cost about a billion dollars, so that’s 1.5 billion transistors per dollar. A lot of the cost is in the chips themselves, but there’s a lot of overhead for network wiring, cabinets, and cooling. It draws 30 MW, as much as a small town.
Now, I’ve worked on about 15 processor chips in my career, and the biggest had a mere 200 million transistors, and drew 4W. My first processor in 1984 had 100 thousand transistors, and drew 10W. This is on an other-worldly scale compared to my systems.
How much is an exa? Consider that one raindrop weighs about one milligram. 1018 raindrops is a billion tonnes of water. That’s about the rainfall on New York City every year. If every transistor in Fugaku were a raindrop, it would take a year to sprinkle them across that huge city.
Or consider that a typical cellphone has 32 GB of storage, and maybe a billion of them are sold each year. That’s 300 exa-transistors. This one machine has 1/200th of all the transistors sold in all the phones.
Who built it? Riken is the leading scientific institution in Japan. It counts four Nobelists among its associates, it isolated the tastes of green tea and umami, it worked on Japan’s atomic bomb during WW II (and was destroyed because of that), and it discovered element 113 (now known as nihonium) in 2004.
Why did they build it? Supers are primarily used to do physical simulations. This is where the interactions between pieces of something obey well-known laws, but there are just so many of them that the overall behavior is unpredictable. The different parts of the simulation interact constantly, so you need a lot of bandwidth among the millions of processor cores. This is different from other big computing complexes, like Google’s or Amazon’s. Those are dealing with millions of separate, independent tasks, ones that don’t need much communication. This class of machine works on single problems and so needs a different (and much more expensive!) structure.
So what have they done with it? Recent publications from the Fugaku group include:
- Discovering through simulation that the glycan molecules on the spike proteins of the COVID-19 virus are critical for infecting cells
- Using machine learning to predict exactly what will flood when tsunamis hit. The Japanese in particular are very, very interested in this.
- Doing the world’s highest resolution weather simulation, with a cell size of 3.5 km across the entire planet.
What’s next? The fastest machine on the Top 500 list has been doubling in speed every 15 months. That’s a factor of 1000 in 13 years. If the overall size of the machine scales with its speed, and it should, then the first zetta-transistor (1021) machine will be in about 2033. There’s no telling who will build it, but IBM has been high up on the list for the list’s entire 30-year history. The first yotta-transistor (1024) machine would come in 2046, just in time for the climate catastrophe to be in full swing, and in full need of simulation. After that, the metric system runs out of prefixes!