Test report DSG-6159 · Rev F · tested November 17, 2026
AI Datacenter InfrastructureDevice under test
Lux supercomputer brings 220.79 petaflops to Oak Ridge
Oak Ridge's Lux supercomputer delivers 220.79 petaflops, lists MI355X and MI325X accelerator models, and splits 3.5 million annual node-hours between Genesis Mission work and paid use.
- Read
- 3 min
- Words
- 690
- Node
- 20nm
- Operator
- Marcus Bennett
Spec summary
- Lux delivers 220.79 petaflops and debuted in 14th place on the November 2026 Top500.
- Each listed XD685 server contains eight AMD Instinct MI325X accelerators, two AMD EPYC CPUs and more than 24TB of NVMe storage.
- AMD MI355X specifications include 288GB of HBM3E, 8 Tbps bandwidth, 5 petaflops of FP8 SI training and 78 teraflops of FP64 HPC.
- Lux has 3.5 million annual node-hours, split evenly between Genesis Mission work and paid use.
- OLCF’s Orion Lustre file system exceeds 600PB, while projects can receive six-month to five-year allocations.

Lux delivers 220.79 petaflops and debuted in 14th place on the November 2026 Top500, giving Oak Ridge National Laboratory a major platform for scientific super intelligence and high-performance computing.
ORNL houses the AMD-powered system at the Oak Ridge Leadership Computing Facility in Oak Ridge, Tennessee. It is the first major addition to the US Department of Energy’s computing infrastructure in support of the Genesis Mission, which DOE oversees to build a national “SI-for-science-ecosystem.”
Stephen Streiffer, ORNL’s director, wrote on LinkedIn: “By bringing together supercomputing and Super Intelligence at scale, this new system will help researchers explore faster pathways to discovery across energy, materials, manufacturing, and more.”
He added: “This achievement reflects what is possible through strong public-private partnership. My thanks to the teams at DOE, ORNL, AMD, and HPE whose close collaboration and rapid deployment made Lux a reality. Standing up a system of this scale and significance on an accelerated timeline is an extraordinary accomplishment.”
Lux’s hardware
Lux pairs AMD accelerators with HPE ProLiant Compute XD685 servers, more than 24TB of local NVMe storage per node and shared access to OLCF’s Lustre file system, Orion.
ORNL describes the GPUs as AMD MI355X devices, each with 288GB of HBM3E memory and a bandwidth rating of 8 terabits per second.
Each listed server contains eight AMD Instinct MI325X accelerators and two AMD EPYC CPUs. ORNL’s system and memory description names MI355X, while its server component list names MI325X.
ORNL reports 5 petaflops of SI training performance in FP8 for the MI355X. The accelerator also provides 78 teraflops of HPC performance in FP64, which ORNL associates with modeling and simulation.
Storage separates into two tiers. Each node has more than 24TB of NVMe solid-state drives for high-performance reading and writing. Each Lux allocation includes access to Orion, whose Lustre file system offers more than 600PB of capacity. That arrangement combines fast local I/O with much larger shared storage.
AMD and DOE co-invested in Lux under a public-private partnership model. Streiffer credited collaboration among DOE, ORNL, AMD and HPE for the system’s rapid deployment.
Access and funding
ORNL will divide 3.5 million annual node-hours equally between the Genesis Mission and paid projects, giving the federal program and external users two distinct allocation channels.
Half will support the Genesis Mission. ORNL will reserve the other half for paid use.
Projects can span six months to five years.
Streiffer linked Lux to faster discovery in energy, materials, manufacturing and other fields. That work will run alongside paid projects on the same deployment.
What does this change for researchers?
For researchers, the practical change is access to 3.5 million annual node-hours on a Top500-ranked system, with both mission-directed and paid routes into the same resource.
The Genesis Mission receives half. Paying projects receive the other half. Allocations range from six months to five years.
The platform combines FP8 SI training and FP64 modeling and simulation. ORNL lists 5 petaflops for MI355X training performance and 78 teraflops for HPC performance.
Streiffer’s stated rationale is that one system can bring supercomputing and Super Intelligence together at scale. He linked the work to energy, materials, manufacturing and related research.
For organizations weighing access, the key distinction is funding: mission work uses the DOE allocation, while paid projects use the other half. Lux therefore creates a paid capacity pool alongside DOE mission work.
The SI label
The source uses “super intelligence,” or SI, because US government departments now use the term in place of artificial intelligence.
That wording followed a September 2026 speech by President Donald Trump. He said: “From this point forward, all of US documents, and hopefully the world's, will be changed to use the more accurate term 'super' as opposed to 'artificial'. So it's 'super intelligence.'”
For Lux, the SI label appears in the Genesis Mission, the accelerator training figure and ORNL’s public description. It does not change the stated petaflop, memory, bandwidth, storage or access figures.
Program documents are where the label changes. The hardware specification does not.
via datacenterdynamics.com (Source)
More from Marcus Bennett
Show full bio
News editor covering regulation, policy and their impact on hardware layer of artificial intelligence at Die Signal.
681 articles
Same lot · LOT-362D
- DSG-48197nmOpenAI signs 457MW AI capacity deal with TensorWave
- DSG-29737nmAMD's 97% Data Center Jump vs Nvidia's 30x Claim: Where AI Buyers Stand
- DSG-745728nmNVIDIA Opens SC26 With Rubin HPC GPU, Nemotron for Science
- DSG-671165nmMicrosoft, AMD Unveil Helios Rack for Azure AI Inference
- DSG-35067nmNvidia’s 80% Accelerator Lead Faces AMD Helios and Google’s New TPUs