Nuclear Fusion HPC: A Pc Weekly Downtime Add podcast

The UK Atomic Vitality Authority (UKAEA) is putting in a brand new supercomputer named Dawn, which is particularly designed to help nuclear fusion analysis. Based mostly on AMD Epyc and Intuition-powered Dell {hardware}, Dawn achieves roughly 6 exaflops at 
at 8-bit precision. Meaning it’s able to working six quintillion floating level operations per second.

Whereas it isn’t as highly effective as Bristol’s Isambard-AI supercomputer, which went stay in June 2025, Dawn is being positioned as probably the most highly effective laptop of its sort particularly designed for nuclear fusion analysis.

“Dawn successfully permits us to tackle a moonshot-like downside much more cheaply, cut back danger and speed up the time of supply of economic fusion units,” Rob Akers, UKAEA director for computing programmes, instructed Pc Weekly.

Constructed on AMD MI 355X graphics processor items (GPUs), Dawn is being optimised for the intersection of AI and high-performance computing (HPC).

At 64-bit precision, Akers says Dawn delivers round 50 petaflops (50 trillion floating level operations per second), which suggests it might run high-fidelity simulations. In line with Akers, the mixed efficiency of 8-bit and 64-bit precision obtainable on Dawn makes it extremely appropriate for AI functions and inference duties. He says: “It’s an extremely highly effective machine, tuned to do some very attention-grabbing work.”

Eight bits was cutting-edge within the heyday of residence computing within the late Seventies and early Eighties. However it’s now a determine that may be very related in right this moment’s excessive efficiency computing. When requested why the 8-bit precision determine issues, given its 64-bit precision determine of fifty petaflops, Akers says: “Excessive efficiency computing floating level precision isn’t the most effective metric for an AI machine.”

As a substitute, an AI machine’s energy is a measure of its potential to course of tokens akin to learn particular person characters in a textual content file to extract that means as Akers explains: “The attention-grabbing factor is that 8-bit precision has turn out to be an extremely highly effective a part of the computing panorama now due to giant language fashions (LLMs). We’re going to be doing work in that house, constructing very bespoke fashions that may ingest textual content  akin to doc archives which have been collected over many, many a long time and switch that into helpful data and data.”

He says Dawn will use AI to assist remedy complicated engineering issues in an economical means. “With out AI, the issues we are attempting to unravel can be intractable,” he says. It is going to be used to create surrogate fashions that may work on a workstation and be capable of  simplify high-fidelity simulations, permitting them to run effectively on smaller methods.

For example, digital twins working on Dawn  shall be used to simulate multi-scale physics and matched methods, which Akers says, means nuclear fusion researchers will in a position to handle emergent behaviours and “black swan” occasions in fusion energy vegetation. “Digital twins give us the flexibility to simulate very high-fidelity fashions, and really importantly, strongly coupled fashions,” he says. “In a fusion energy plant, there are many completely different bodily mechanisms that couple the plant collectively—every thing from structural forces as a result of gravity, but in addition as a result of electromagnetism after which there’s loads of warmth stream and radiation stream throughout the system. Every thing’s coupled collectively.”

Through the use of Dawn to run digital twin simulations, the UKAEA goals to enhance confidence in designs and allow breakthroughs in equipment improvement for nuclear fusion.