Japan Plans Zettaflop Supercomputer by 2030

by

Japan plans to spend about $761 million to build a Zettaflop supercomputer. It is called Fugaku Next. It will be built by Japanese companies RIKEN and Fujitsu, which were both involved in the construction of Fugaku.

The AI training clusters with lots of GPUs can get to a peak of a Zettaflop but this is for FP8 or FP4 calculations.

Intel helped Japan to do a feasibility study.

The computational node of the Fugaku Next supercomputer will have peak performance of several hundred FP64 TFLOPS for double-precision computations, around 50 FP16 PFLOPS for AI-oriented half-precision calculations, and approximately 100 PFLOPS for AI-oriented 8-bit precision calculations, with memory bandwidths reaching several hundred TB/s using HBM-type memory.

Floating-point operations per second (FLOPS) is used to measure how fast computers can solve problems — where one floating-point operation is a single calculation. A supercomputer with a speed of 1 zettaFLOPS could make one sextillion (1 followed by 21 zeros) calculations per second. Today’s most powerful supercomputers have only just broken the exaFLOPS barrier, meaning they can make just over one quintillion (1 followed by 18 zeros).

MEXT has set aside around ¥4.2 billion ($29 million) for the first year of the project but could allocate up to ¥110 billion ($761 million) throughout the project, which is scheduled to be completed by 2030

7 thoughts on “Japan Plans Zettaflop Supercomputer by 2030”

  1. This is hard for me to explain, because I’m not a “computer guy”. But I’ll try. Your right, current digital technology is shockingly limited. Please explain what “mantissa” means. I do not understand that word. Hey, I don’t understand something, I keep asking questions. I will keep “asking questions”, until one, I get an answer, that (far as I know), answers that question. Or two, the “answer” just makes me “feel good”. I can’t tell you how wonderful or important this last point is. Very often “intangible”, but in the end, it’s the difference between doing what matters, and doing what matters to each of us, one man or woman at a time.

    • By the way, I have enough trouble understanding a billion, let alone a trillion. A “zettaflop”? Just because I know others do not understand this “volume or speed”, does not make me “feel better” Being with a bunch of other people who do not understand something, hardly makes me “feel good”. It’s quite scary, when someone doesn’t understand something, and doesn’t care. Such technology may soon enough effect you, or others. Wouldn’t you like to know how? I would.

  2. FP4?! How does that work, 1 bit for sign, 1 bit for exponent, and 2 bits for mantissa? Even 8 bits is barely functional…

    • This is hard for me to explain, because I’m not a “computer guy”. But I’ll try. Your right, current digital technology is shockingly limited. Please explain what “mantissa” means. I do not understand that word. Hey, I don’t understand something, I keep asking questions. I will keep “asking questions”, until one, I get an answer, that (far as I know), answers that question. Or two, the “answer” just makes me “feel good”. I can’t tell you how wonderful or important this last point is. Very often “intangible”, but in the end, it’s the difference between doing what matters, and doing what matters to each of us, one man or woman at a time.

  3. Not very fast though.. That’s 5 years. I know that it’s not apples to apples, but xAI managed to set up their cluster in, what, 4 months? The risk is that this computer will be obsolete by the time it’s online.

    • Back in the 1980’s my first (and only) company developed an optical fingerprint imaging technology. We had to scan prints, send them to the FBI AFIS (automated fingerprint identification system), and send ID back to source, like a police station, in close to real time. In those days, this was “rather cool”. Then, the fastest commercially available integrated circuit was the 8080 or 8888 Intel chip. By todays standards, it was shockingly slow. So we used math, in the form of a fractal, to deal with that problem. Most people think of fractals as pretty pictures. Which they are. But oh, they are so much more.

      The information in a fractal is the same, no matter the volume. Expand or contract a fractal, it’s “innate nature, or contained information”, does not change. This may be a great way to differentiate between data, and information. Data is content, information is context. Back then, with the available technology, we had to be creative. Oh yes we we’re. God, was it fun. My point? We can bypass the limitations of our current technology if we can imagine how. I (and you) can imagine quite a lot.

Comments are closed.