UKAEA launches Sunrise supercomputer to speed up fusion research with AI and high-fidelity simulations

The UK Atomic Energy Authority has begun installing a purpose-built supercomputer, Sunrise, to accelerate research into commercial nuclear fusion by combining high-performance simulation with large-scale artificial intelligence workloads. According to Computer Weekly, the system is built on AMD Epyc processors and AMD Instinct accelerators housed in Dell hardware and is optimised around AMD MI 355X GPUs. Rob Akers, UKAEA director for computing programmes, told Computer Weekly that Sunrise delivers roughly 6 exaflops at 8-bit precision, the equivalent of six quintillion operations per second, and about 50 petaflops at 64-bit precision for high-fidelity physics modelling.

UKAEA frames Sunrise as a targeted platform for the specialised demands of fusion engineering rather than a general-purpose national flagship. “Sunrise effectively allows us to take on a moonshot-like problem a lot more cost effectively, reduce risk and accelerate the time of delivery of commercial fusion devices,” Akers said to Computer Weekly. The authority argues that pairing very low-precision throughput for AI with conventional double-precision capability for detailed simulations enables new workflows: large language models and other AI tools can be trained and deployed on token-processing tasks while the same facility runs high-accuracy physics codes.

Akers explained the operational rationale for the dual-precision approach in the Computer Weekly interview: “High performance computing floating point precision is not the best metric for an AI machine.” He added that 8-bit arithmetic has resurged in relevance because of advances in large language models and their use in extracting meaning from long archives: “The interesting thing is that 8-bit precision has become an incredibly powerful part of the computing landscape now because of large language models (LLMs). We are going to be doing work in that space, building very bespoke models that will ingest text such as document archives that have been collected over many, many decades and turn that into useful information and knowledge.”

Sunrise is being positioned to underpin surrogate modelling and digital-twin workstreams that compress expensive, high-fidelity simulations into lighter-weight models suitable for routine design and optimisation. “Without AI, the problems we are trying to solve would be intractable,” Akers said, speaking to Computer Weekly. He described how digital twins running on Sunrise will represent strongly coupled, multi-physics behaviour, structural, electromagnetic, thermal and radiative interactions, so researchers can probe emergent failure modes and rare “black swan” events in plant designs and build greater confidence before committing to hardware manufacture.

The UKAEA’s strategy aligns with broader trends in the UK research ecosystem toward AI–HPC convergence but stops short of claiming national primacy. The University of Cambridge, together with Intel and Dell Technologies, has deployed Dawn at the Cambridge Open Zettascale Lab, which has been promoted as the UK’s fastest AI supercomputer and is intended to support work across fusion energy, healthcare and climate modelling. Cambridge and industry partners describe Dawn as a co-designed system built around 4th Gen Intel Xeon Scalable processors and Intel Data Center GPU Max accelerators, optimised for dense AI training workloads.

Industry reporting and vendor announcements indicate the two machines pursue different trade-offs: Sunrise targets fusion-specific mixed workloads with AMD accelerators and a balance of extreme low-precision throughput and traditional double-precision capability for fidelity; Dawn targets maxed-out AI training throughput with Intel CPUs and Data Center GPUs across a larger GPU count and greater aggregate AI memory bandwidth. According to Data Center Dynamics and Intel’s newsroom, Dawn comprises hundreds of nodes and more than a thousand Intel GPUs, reflecting a design point aimed at general-purpose national-level AI research.

For industrial decarbonisation stakeholders and engineering teams, Sunrise’s arrival matters not merely as a benchmark in peak arithmetic, but as an applied toolchain for reducing development cycles and de-risking capital-intensive hardware. UKAEA expects the platform to produce surrogate models that engineers can run on smaller workstations to iterate designs, to speed up design-of-experiment campaigns, and to allow more exhaustive exploration of failure envelopes before factory-scale builds.

The UKAEA’s public remarks emphasise practical outcomes rather than raw rankings. By combining 8-bit AI throughput with 64-bit simulation power and by embedding digital-twin development into fusion R&D, Sunrise is intended to shorten the path from experimental device to commercially viable plant while lowering programme risk. As the nation’s supercomputing landscape fills with purpose-built installations such as Dawn and Sunrise, fusion developers and industrial decarbonisation planners will need to match application requirements, precision, memory bandwidth, and software co-design, to the increasingly specialised hardware ecosystems being deployed.