HPE's Lux AI cluster has debuted at No. 14 on the TOP500 with 220.79 measured petaflops, built for Oak Ridge National Laboratory on AMD Instinct MI355X GPUs. For UK buyers evaluating sovereign AI infrastructure and energy-constrained deployments, it is the first large-scale MI355X cluster with a TOP500-benchmarked result to reference against vendor claims.
View the data behind this chart
| Measured HPL (TOP500) | Rated performance | |
|---|---|---|
| Petaflops | 220.79 PF | 323.9 PF |
| 0 PF | 0 PF |
What actually launched
HPE confirmed that Lux, the AI cluster it built with AMD for Oak Ridge National Laboratory, recorded 220.79 petaflops of reported HPL result to take 14th place on the November 2026 TOP500 list; HPE says Lux ranked ahead of several systems designed for traditional supercomputing workloads. It is the first dedicated AI system to come online for the US Department of Energy's Genesis Mission, with half of its 3.5 million annual node-hours reserved for that programme and the rest available for proprietary use — a split approximately equivalent to 400 nodes running continuously for a year.
For UK infrastructure buyers, the value is not that Lux exists — it is that a major vendor pairing has now published a real, published HPL benchmark result for AMD's current-generation accelerator at scale, rather than relying solely on peak-performance marketing figures.
Inside the build: compute, cooling, networking
Lux runs on direct liquid-cooled HPE ProLiant Compute XD685 systems combining AMD Instinct MI355X GPUs, AMD EPYC CPUs and AMD Pensando DPUs. Tom's Hardware reports more than 4,000 GPUs across more than 500 nodes; HPE and AMD have not published an official count. At least roughly 1.15PB of aggregate HBM3E memory would follow if the reported 4,000-plus-GPU figure is accurate, with each GPU carrying 288GB. The network layer uses HPE Juniper Networking QFX switches on 800G Ethernet with congestion management intended to lift GPU utilisation, managed through HPE Networking Data Center Director — HPE describes Lux as its first deployment of HPE Networking products at a DOE national laboratory.
Buyers scoping similar builds should treat the cooling and fabric choices as closely linked to the GPU selection itself. Our explainer on direct-to-chip liquid cooling covers why MI355X-class thermal density can make liquid cooling highly attractive or necessary at high rack densities, though air-cooled MI355X options do exist according to AMD documentation. Our breakdown of 800G Ethernet fabric designs is directly relevant given Lux uses an 800G Ethernet back-end network.
Reading the benchmark numbers correctly
The 220.79-petaflop HPL result is a conventional HPC benchmark result, not an AI training or inference throughput figure — The Register's coverage of the same list makes this distinction explicit. The Next Platform separately cites Lux's theoretical rated performance at 323.9 petaflops, while Tom's Hardware and related reporting on the same MI355X platform cite reported platform peak figures as high as 20 exaflops for FP8 training and 40 exaflops for FP4 inference. These numbers describe different things and are not comparable on a single chart.
For procurement teams, the lesson is to demand workload-specific figures rather than a single headline number. Dell reports that an eight-GPU MI355X PowerEdge XE9785L system achieved 70,610.8 tokens per second on its MLPerf Inference v6.0 large-language-model workload, and separately recorded 78,603.5 queries per second in a distinct server-mode test — two separate measurements that should not be treated as interchangeable, and which are more operationally useful for inference-heavy deployments than HPL alone. Teams modelling their own cluster requirements can run these variables through an AI GPU cluster calculator before committing capital, and compare the platform generally against alternatives using our AMD Instinct enterprise deployment analysis.

Power and energy planning
The MI355X carries a board power of up to 1,400W typical board power/TBP and delivers 78.6 TFLOPS of FP64 performance, compared with the cited 750W MI300X board-power figure from AMD and Dell documentation. 4,000 GPUs at 1,400W each would represent 5.6MW of accelerator nameplate power; a reported 4,000-plus-GPU installation would be higher, before CPUs, networking, memory and cooling overheads are added — a figure UK buyers should weigh heavily given grid connection constraints and electricity costs.
In Dell's tests, MI355X runtime power ran between 943W and 1,256W, below the 1,400W typical board power figure, and Dell's MLPerf submission separately reported that power-capping an MI355X configuration from 1,400W down to 1,000W produced only a modest performance reduction. These are Dell platform/workload results, not measurements taken from Lux itself. That kind of headroom matters directly for sites with constrained rack power or fixed energy budgets, and should be scoped early using proper server configuration planning rather than assumed at nameplate figures.
Sovereign AI and public-sector relevance
AMD's Thomas Zacharia linked Lux to what he described as an open, sovereign AI stack usable by researchers nationally, and HPE has made a $37 million in-kind investment to support Genesis Mission research projects. Lux is a US Department of Energy deployment, not a UK sovereign-cloud project, but it is a concrete, benchmarked reference architecture for public-sector and regulated buyers assessing domestically controlled AI capacity rather than renting from hyperscale providers.
UK organisations weighing the build-versus-rent question on sovereign capability can use Lux's documented specification — liquid-cooled nodes, 800G Ethernet, multi-tenant cloud-style access — as a concrete benchmark when working through our sovereign AI infrastructure guidance, and when comparing it against data-centre-class alternatives such as NVIDIA DGX systems.
View the data behind this chart
| MI355X | MI300X | Why it matters | |
|---|---|---|---|
| Board power (TDP) | 1,400W | 750W | Higher density, higher cooling load |
| HBM per GPU | 288GB | Lower | More headroom for large models |
| Runtime power (tested) | 943–1,256W | Lower baseline | Below TDP in real workloads |
| Power-capped option | 1,000W cap, modest loss | Not applicable | Useful for constrained grid capacity |
What UK buyers should actually benchmark
AMD's share of new TOP500 entries exceeded 40% on this list according to Tom's Hardware, a visibility signal for the platform rather than proof of equivalent commercial availability or support maturity in the UK market. Treat Lux as one well-documented data point, not a template to copy outright.
Before any procurement decision, pull together HPL or equivalent HPC scores, model-specific tokens-per-second figures, useful work per kilowatt-hour, rack density, liquid-cooling requirements, HBM capacity, ROCm software compatibility and total delivered cost — the single TOP500 ranking is not sufficient evidence on its own for sovereign-AI or energy-critical purchasing decisions. Our wider AI servers data study sets out how to weigh these variables against each other systematically.
- 01StorageReview — HPE Lux AI Cluster Debuts at No. 14 on the TOP500 With 220.79 Petaflops on AMD Instinct MI355X GPUs · 16 November 2026
- 02HPCwire — HPE-Built 'Lux' AI Cluster Debuts at No. 14 on the TOP500 · 17 November 2026
- 03The Register — China's grip on TOP500 remains uncontested as US chases AI, HPC, fusion · 16 November 2026
- 04Tom's Hardware — AMD grabs over 40 percent of new TOP500 list for supercomputers · 16 November 2026
- 05The Next Platform — Supercomputer installs take a dive in the second half of 2026 · 16 November 2026
- 06Dell — The TCO Titan: MI355X vs MI300X on Dell PowerEdge Servers · 16 November 2026
- 07Dell — MLPerf Inference v6.0: Dell Showcases Performance with AMD Instinct MI355X GPUs · 16 November 2026
- 08storagereview.com
