Groundbreaking NVIDIA experiment sets computer simulation record

A ground-breaking physics experiment led by the University of Essex and accelerated computing and AI leader NVIDIA has set a new record in computer simulations, unlocking possibilities for next-generation tech.

Dr Nikolaos Fytas, from Essex’s School of Mathematics, Statistics and Actuarial Science, teamed up with NVIDIA engineers to run the largest simulation ever performed in statistical physics.

The pioneering work could help the development of a new wave of displays, advanced magnetic materials, and deeper insights into the building blocks of matter.

Their work made the first practical observation of the “thermodynamic limit,” a long-sought milestone in physics that explains how the properties of matter emerge when systems become extremely large.

'Uncover nature's secrets'

“This marks a key scientific achievement,” said Dr Fytas.

“It gives us a clearer picture of how size and simulation time are linked and reveals universal patterns in the way materials behave.

“None of this would have been possible without close collaboration between physicists and computer scientists, and it shows the incredible power of modern computing to uncover nature’s secrets.”

The team focused on two classic models in physics, the Ising model and the Blume-Capel model, which mimic how tiny magnetic particles flip and interact.

This helps scientists understand how matter changes state, such as when metals lose their magnetism when heated.

'New benchmark'

Using NVIDIA’s latest GB200 NVL72 rack-scale architecture and advanced memory systems, they developed a code capable of simulating up to 70 trillion interacting particles, far beyond what exists in a speck of visible matter.

The fabric memory capabilities of NVIDIA NVL72-enabled systems made such large-scale simulations possible, allowing all GPUs to access each other’s memory as a unified address space over high-bandwidth, low-latency NVLink connections.

This eliminated the need for explicit data transfers, simplified implementation, and delivered near-perfect strong scaling.

The system hit a record-breaking speed of almost 115,000 lattice updates per nanosecond.

This set a new benchmark for these types of simulations.

The research has been published in Physical Review Research and Computer Physics Communications.