US to build a supercomputer ten times more powerful than today’s systems

The U.S. Department of Energy запросило information from suppliers to build a supercomputer with a capacity of more than 10 exaflops. The installation is planned to be launched by 2030.

According to the document, the department is interested in deploying one or more supercomputers capable of solving high-precision tasks five to ten times faster than current systems.

By 2030 they expect to assemble a machine with 10–20 exaflops. After 2030, the plan is to boost performance to 100 exaflops, to be achieved “through hardware and software acceleration”.

Power consumption is expected to be 20–60 MW.

The department also said it aims to move away from “monolithic installations” toward modular systems. This would accelerate the adoption of innovations in hardware and software.

As a result, the component refresh cycle for the supercomputer could fall from four- to five-year intervals to 12–24 months, the department said.

The information the Department of Energy is seeking from suppliers covers prospective directions for 2025–2030:

• types of processors, memory and storage;
• interconnect options;
• system-on-a-chip combinations;
• promising process technologies;
• expectations regarding bandwidth;
• potential node configuration, and more.

The document also notes that future installations will require a software stack capable of handling a wide range of tasks in large-scale modelling, machine learning and data analytics. They must be fault-tolerant to minimise manual labour.

Vendors must submit their proposals by the end of July 2022.

Earlier in May, the American supercomputer Frontier took top spot in the Top500 rankings. It was the first installation to reach a peak of 1.1 exaflops in the Linmark test.

In June, the American startup Cerebras trained the “largest AI model” on a single device, setting a world record.

In May 2021, scientists unveiled an AI supercomputer that will help construct the largest-ever 3D map of the visible universe to study dark energy.

Subscribe to ForkLog news on Telegram: ForkLog AI — all the news from the world of AI!