AMD GPUs to train Arena Physica’s next electromagnetism model on Oak Ridge supercomputer

Arena Physica plans to train future versions of its Heaviside electromagnetism model on AMD-powered Lux, Oak Ridge National Laboratory’s supercomputer. It is a substantial AI-computing project, but the training is still ahead: no Lux-trained results or performance gains have been reported yet.

AMD Watch analysis

What happened

Arena Physica announced the collaboration on 12 November, according to Unite.AI. The work builds on Heaviside-1, a 350-million-parameter model released in September. Arena Physica says that version was trained on 250,000 three-dimensional electromagnetic simulations, containing 500 billion field samples.

Lux is listed by Oak Ridge’s Leadership Computing Facility as having more than 4,000 AMD MI355X GPUs, 1.1 petabytes of high-bandwidth memory and 20 exaFLOPS of FP8 AI-training performance. Those are specifications for the system, not a benchmark of Heaviside running on it. Arena Physica says it plans to share results as training progresses.

The company’s stated ambitions for later model generations include handling wider frequency ranges, more complex geometries and larger systems. Potential applications include antennas for mobile and satellite communications, automotive radio-frequency sensing, and semiconductor packaging and signal integrity.

Why it matters

Electromagnetism is difficult to model because small details in a design can affect the behaviour of the whole structure. The project pairs a model trained on simulation data with a large compute system, aiming to extend what it can handle. That makes this a concrete AI-compute collaboration, rather than an AMD logo wandering past a generic technology story.

If the approach works, faster predictions could help engineers explore designs that would otherwise take more time or computing resources with traditional solvers. Arena Physica also describes autonomous inverse design as a goal. Those are development aims, not demonstrated results from the Lux collaboration.

Our read

This is a promising use of AI in a specialised engineering field, with unusually useful detail about both the training data and the machine doing the work. The next evidence to look for is not another impressive hardware specification, but results: what the model can predict, how accurately, and against which conventional methods. For now, the collaboration is a plan with serious hardware behind it, not proof that the hard part is solved.

What to watch

• Arena Physica’s forthcoming results from training on Lux.
• Whether later Heaviside versions cover the larger geometries and wider frequencies the company has outlined.
• How any claimed speed or cost gains compare with traditional electromagnetic solvers.

Discussion spark: If this model delivers useful predictions faster than traditional electromagnetic solvers, which should matter most to engineers: accuracy, cost, or the range of designs it can handle?

Independent WittyWires tracker for public updates about AMD. Not affiliated with or endorsed by AMD; this is not an official account.