Lux, a supercomputer at Oak Ridge National Laboratory (ORNL) in the United States, began operating for scientific researchers on November 16, 2026. Equipped with 4,032 GPUs, it will serve as a computing platform supporting research that combines AI, simulation, and experimental data. Lux was built through a joint investment by AMD and the U.S. Department of Energy (DOE), and roughly half of its 3.5 million node-hours of annual computing capacity will be allocated to the DOE's Genesis Mission. It is expected to be used for designing fusion power plants, developing materials, forecasting water resources, and more. In AMD's announcement, researchers and organizations that want to use it are advised to consult ORNL's computing facility about their computing needs and how to access the system.
4,032 GPUs in a platform that integrates AI and simulation
Lux was built through a public-private investment by AMD and the DOE, and Hewlett Packard Enterprise (HPE) constructed it. It combines AMD's EPYC CPUs, Instinct MI355X GPUs, Pensando networking products, and the ROCm software stack.
According to the Lux user guide, the full system consists of 504 nodes. A node is a unit of computing hardware that includes CPUs, GPUs, memory, and other components. Each node has two 64-core EPYC 9575F processors, eight Instinct MI355X GPUs, and 3 TB of DDR5 memory. Each GPU also has 288 GB of high-speed HBM3E memory.
Lux divides its computing resources into two partitions: one managed by Slurm, a job management system for HPC, and one managed by Kubernetes, which manages the execution of containerized applications. The number of nodes assigned to each depends on how the system is partitioned. According to ORNL's explanation, the system is intended for use that combines conventional modeling and simulation with AI training and inference, as well as the management of research workflows by AI agents.
Infrastructure for storing and sharing research data is also in place. Lux connects to Orion, a shared file system based on Lustre and HPE ClusterStor. Frontier, another supercomputer at the same facility, along with data transfer nodes and Riker, also uses Orion, and the available namespace capacity reaches 679 PB. However, Lux's compute allocations are managed separately from those for Frontier and other systems. Even though the file system is shared, eligibility to use each machine and access rights to data follow the conditions that apply to each.
As for performance, according to AMD's announcement, Lux ranked 14th on the November 2026 edition of the TOP500 supercomputer list, recording 220.79 petaflops. This is a computing performance figure measured with the HPL benchmark, and it does not directly indicate AI model training speed or inference throughput.
How will 3.5 million node-hours a year be allocated?
According to ORNL, of the 3.5 million node-hours Lux provides each year, about half goes to the DOE's Genesis Mission and the other half to AMD. ORNL describes this as a model in which the public and private sectors share the costs of building and operating the computing infrastructure. A node-hour is a unit of computing usage based on using one entire node for one hour.
Researchers, however, do not always occupy an entire node. In Lux's Slurm environment, multiple users can share the same node. According to the user guide, requesting one GPU allocates 16 CPU cores and 375 GB of host memory, so the job counts as using one-eighth of a node. The 375 GB is the host-side memory allocation and is separate from the 288 GB of HBM3E on each GPU.
For Slurm batch jobs, running on one GPU for 8 hours and running on 8 GPUs for 1 hour each consume one node-hour of allocation. Using the calculation method described in the user guide, node-hours consumed can be expressed as follows.
\[ \text{Node-hours consumed}=0.125\times\text{GPUs requested}\times\text{runtime (hours)} \]
For example, requesting one GPU for 8 hours gives , and requesting eight GPUs for 1 hour gives . Both are counted as one node-hour, but that does not mean the performance or research results are the same.
This method is based on the computing resources that a job's request makes unavailable to other users and the time the job was in a running state. It does not calculate usage based on the proportion of time the GPUs were actually performing computations. The example also applies to Slurm batch jobs; Kubernetes and reserved time slots have different conditions.
It is therefore necessary to distinguish between the 4,032 GPUs installed in the whole system, the number of GPUs an individual job reserves, and the node-hours consumed by running it. For researchers and organizations that want to use Lux, ORNL's Oak Ridge Leadership Computing Facility (OLCF) accepts inquiries about the computing resources needed and the allocation programs that may apply.
From fusion power plant design to materials development and water forecasting
One research effort using Lux's computing resources is AI-driven fusion power plant design. According to plans ORNL described, the aim is to develop an AI platform that integrates HPC tools with robotic system design. Research is also planned to build foundation models specialized for materials development and advanced manufacturing.
Earth system prediction is another important area. For regions involving the Tennessee Valley Authority (TVA), plans call for developing technology to forecast water availability, river flow, and water temperature from several weeks to several years ahead. The goal is to use these forecasts in operating hydroelectric and thermal power plants. Although the subjects differ from fusion and materials development, all of these efforts seek to incorporate AI into conventional scientific research to improve methods of research and prediction.
ORNL says that each research team participating in the Genesis Mission will design and demonstrate methods that integrate AI and scientific research, and will rigorously evaluate whether they lead to faster scientific discovery, better predictions, and more advanced experiments. With Lux now in operation, the large-scale computing resources these studies require are available. What methods each team actually establishes, and how much they improve on conventional predictions and experiments, will be the basis for judging the results ahead.
