Home ComputersAMD Lux Supercomputer Opens With 4,032 Instinct MI355X GPUs

AMD Lux Supercomputer Opens With 4,032 Instinct MI355X GPUs

by Warren
17 views
AMD Lux supercomputer system at Oak Ridge National Laboratory

AMD’s Lux supercomputer is now open to scientific users at Oak Ridge National Laboratory, bringing a large pool of Instinct MI355X accelerators to the US Department of Energy’s Genesis Mission.

Built by HPE, the system combines AMD EPYC server processors, Instinct GPUs, Pensando networking and the ROCm software stack. AMD describes Lux as the first dedicated US AI factory for the Genesis Mission, where it will support work spanning fusion energy, advanced materials, manufacturing and Earth-system modelling.

Lux supercomputer server aisle at Oak Ridge National Laboratory
Lux is housed at Oak Ridge National Laboratory in Tennessee. Image: AMD/Oak Ridge National Laboratory

More than 4,000 Instinct MI355X GPUs

Oak Ridge’s technical documentation gives Lux a total of 504 compute nodes and 4,032 AMD Instinct MI355X GPUs. Each node contains two 64-core AMD EPYC 9575F processors, 3TB of DDR5 memory and eight MI355X GPUs. Every accelerator provides 288GB of HBM3E memory, giving researchers substantial local memory capacity for large AI models and scientific data sets.

The November 2026 TOP500 list places Lux at number 14, with 220.79 petaflops of measured HPL performance. It joins El Capitan, Frontier, HPC7 and HPC6 among the AMD-powered systems near the top of the global rankings. AMD says its technology now appears in 202 of the 500 listed machines.

AMD Lux supercomputer compute section with Genesis Mission branding
The system combines EPYC processors, Instinct accelerators and Pensando networking. Image: AMD/Oak Ridge National Laboratory

One network for AI training and simulation

Lux uses AMD Pensando Pollara 400 AI NICs to connect the accelerators over a common Ethernet fabric. Each GPU is paired with a 400Gbps network interface, allowing the same infrastructure to handle large AI training jobs and the MPI communication used by traditional scientific simulations.

AMD says a 123-billion-parameter language model reached 4.48 exaflops of FP8 throughput across 2,048 MI355X GPUs in its testing. The company also demonstrated scientific workloads scaling across thousands of accelerators, including a molecular-dynamics simulation of 17.2 billion tungsten atoms.

AMD Pensando AI NIC networking for AI and scientific computing
AMD’s programmable AI NIC platform now supports both AI and HPC workloads. Image: AMD

Access is aimed at scientific research

Lux is not a commercial computer that organisations can buy off the shelf. Researchers can instead approach the Oak Ridge Leadership Computing Facility to discuss access for suitable Genesis Mission projects.

Initial work allocated to the system includes AI-assisted fusion power-plant design, foundation models for materials and advanced manufacturing, and tools for forecasting water availability, streamflow and river temperatures. AMD and Oak Ridge are also working on the American Science Cloud, which is intended to extend advanced AI and high-performance computing resources to other laboratories, universities and mission partners.

No Malaysian access programme or local commercial availability was announced.

You may also like