Category: Industry News · Draft date: 2026-11-22
AMD compute has cracked the TOP500 top 15. On 2026-11-16, at SC26 in Chicago, the 68th TOP500 list landed with Lux — the AI cluster HPE built for Oak Ridge National Laboratory (ORNL) — debuting at No. 14 with 220.79 PFLOPS of measured HPL on AMD Instinct MI355X accelerators. This is not another traditional supercomputer chasing a badge: Lux is the first dedicated AI system online for the U.S. Department of Energy's (DOE) Genesis Mission, with half its machine-hours reserved for national research and the other half at AMD's disposal. A machine designed for AI training and inference casually walked into a ranking that still crowns double-precision champions — and that fact is more interesting than the rank itself.
One line on the list tells you the whole configuration

The TOP500 entry for Lux reads: HPE ProLiant Compute XD685 (direct liquid-cooled), AMD EPYC 9575F (64 cores at 3.3GHz), MI355X, HPE Juniper/AMD Pensando 400G Ethernet, RHEL 9; 1,096,704 total cores, Rmax 220.79 PF, Rpeak 323.90 PF, power not reported.
Neither HPE nor AMD published a node or GPU count, but the listed core count lets you reverse-engineer it (calculated): each XD685 node carries 2 sockets of 64-core EPYC plus 8 MI355X cards (256 CU each), for 2,176 counted cores; 1,096,704 ÷ 2,176 = 504 nodes, meaning 4,032 MI355X GPUs and 1,008 EPYC CPUs. That matches press reports of "4,000+ GPUs across 500+ nodes."
Per-card MI355X specs (AMD official, launched 2025-06-12): 288GB HBM3E, 8TB/s bandwidth, 256 CU, MXFP8 peak of 5 PF and MXFP4 at 10.1 PF. Across 4,032 cards that works out to roughly 20.16 EF (FP8) and 40.72 EF (FP4) of system low-precision peak (calculated), corroborating the "about 20/40 EF" figures in the press. Subtract the double-precision contribution of 1,008 EPYC sockets (about 6.81 PF, calculated) from Rpeak 323.90 PF and the GPU side holds roughly 317 PF of FP64 — about 78.6 TF per card (calculated), or roughly 1/64 of its MXFP8 peak. Lux's muscle is built for low precision; HPL is a side exam it took in passing.
One small discrepancy in the networking record: TOP500 lists 400G Ethernet (the node-side Pensando DPU), while HPE's press release stresses QFX switches supporting 800G — the two figures describe NIC ports versus switch ports (inferred). HPE calls this the first deployment of its Networking products at a DOE national laboratory, managed through Data Center Director for intent-based operations.
What No. 14 really means: LINPACK's yardstick and a historically quiet list

TOP500 ranks HPL — measured Rmax on dense FP64 linear systems — not AI throughput. Lux's HPL efficiency is 220.79/323.90 ≈ 68.2% (calculated), in the same band as El Capitan's 1,809.00/2,821.10 ≈ 64.1% (calculated), which is respectable. But its total FP64 is only about 1/6 of Frontier's (1,353.00 PF, No. 3) at the same lab (calculated: a gap of roughly 6.1×). Pricing an AI machine by its LINPACK rank will systematically misread it.
This edition of the list was unusually quiet: the top 10 is identical to the 2026-06 edition, only 27 systems are new (5.4%) — the lowest turnover ever — yet the No. 500 threshold rose to 2.857 PF (+7.3%), total list HPL reached 19.57 EF (+4.4%), and accelerator-based systems hit 291 (58.2%). The official TOP500 press release cites Epoch AI data to explain where the hardware went: quarterly additions to global AI compute are growing roughly 2.5× per year, cumulative capacity has passed 15 million H100 equivalents, and as of Q4 2025 five U.S. hyperscalers owned about 71% of it. Most new silicon is going into "AI factories" that never enter this list.
Lux is precisely the interface between those two worlds: built like an AI factory (multi-tenant, cloud-style access, dual Slurm and Kubernetes scheduling), yet it showed up to register on TOP500. Ahead of it at No. 13 sits Leonardo (241.20 PF), a traditional HPC machine from the A100 era; behind it at No. 15 is Isambard-AI phase 2 (216.50 PF), a GH200 AI system. Lux wedged between them is the portrait of AI machines pushing into HPC's heartland.
AMD vs. NVIDIA: two strategies in the top 20

Counting the top 20 rows of the list (tallied by this article): 7 AMD-accelerated systems (No. 2/3/6/8/11/14/16), 10 NVIDIA, 1 Intel (No. 4 Aurora), and 2 custom-chip machines (No. 1 LineShine, No. 9 Fugaku). Sum Rmax instead, and AMD's roughly 5,020 PF beats NVIDIA's roughly 3,494 PF (calculated): AMD's head is heavier — El Capitan (MI300A) and Frontier (MI250X), two exascale machines, plus HPC6/HPC7 — while NVIDIA's footprint is broader, blooming across clouds and AI factories (Eagle, ISEG2, Maximus-02 and others).
Nor is Lux the only MI355X deployment: No. 65 Sunrise (UKAEA/Cambridge, 42.32 PF) and No. 71 Zenith (Cambridge, 36.33 PF) are Dell systems pairing EPYC 9575F with MI355X, and Lux is the largest MI355X installation to date. HPE's official blog, meanwhile, notes that 10 of the top 15 systems were built by HPE — the winners include integration and liquid-cooling engineering, not just chip vendors. The crown itself now belongs to Shenzhen's LineShine (2.198 EF, custom LX2), and TOP500 has entered a multipolar era. What AMD proved with MI355X is something else: an open ROCm software stack, standard servers and open Ethernet can also carry leadership-class AI compute.
What research buyers should copy: the model, not the spec sheet
Lux's procurement model deserves more study than its configuration. Public-private cost sharing: the AMD-DOE agreement totals $1 billion and covers two systems (Lux and Discovery, delivering in 2028 with Venice EPYC plus MI430X); HPE added an in-kind investment of $37 million. DOE takes half the machine-hours — 3.5 million node-hours a year, equivalent to about 400 nodes running nonstop (StorageReview's conversion) or roughly 79% annual utilization across the calculated 504 nodes (calculated); AMD takes the other half to mature its software and workloads. The lab contributes site and operations, the vendors contribute capital and hardware, and both sides share the risk.
Buy workflows, not rankings: Genesis Mission projects on Lux run fusion power plant design, a materials "foundation model factory," and TVA-basin water prediction — hybrid AI-plus-simulation tasks. Slurm coexisting with Kubernetes means one machine runs both MPI jobs and containerized inference. For universities and research institutes, acceptance criteria should be throughput on your own hybrid workflows and machine-hour utilization, not a LINPACK peak.
The commodity-hardware route works: XD685 is a standard ProLiant liquid-cooled server — the same chassis also hosts NVIDIA B200 systems (NHN Cloud, No. 22/43) — and the fabric is open Ethernet rather than InfiniBand. For universities, testing agencies and corporate research labs in Guangzhou and the Pearl River Delta building their own AI-plus-simulation platforms, this "standard servers + Ethernet + dual schedulers" route lowers the bar for supply, operations and spares. ROCm proven at leadership scale also weakens the "CUDA or nothing" assumption.
No. 14 is not the end point: after Discovery arrives in 2028, ORNL will field an exascale simulation machine and a leadership AI machine side by side. The next "quiet" TOP500 may well carry more machines like Lux that refuse to stay in their lane.
Sources (URL + date, all verified on 2026-11-22)
- TOP500 68th list (Lux row: cores, Rmax, Rpeak; basis for the top 20 tally): https://www.top500.org/lists/top500/list/2026/11 , published 2026-11-16
- Official TOP500 press release (new-system count, thresholds, Epoch AI data): https://www.top500.org/news/top500-leadership-holds-steady-new-system-turnover-falls-historic-low/ , 2026-11-16
- ORNL/OLCF official release (node-hour split, Slurm + Kubernetes, Genesis project list): https://www.olcf.ornl.gov/2026-11-16/lux-genesis-mission/ , 2026-11-16
- AMD official release (first Genesis Mission AI Factory, ROCm/Pensando): https://newsroom.amd.com/news/amd-powered-lux-supercomputer-opens-scientific-users/ , 2026-11-16
- HPE press release (Business Wire; 800G QFX, in-kind investment; full text verified via ayondo syndication): https://ayondo.com/en/bwr/US42824C1099/hewlett-packard-enterprise/hpe-built-lux-ai-cluster-debuts-at-no-14-on-the-top500-20261116076971r1 , 2026-11-16
- HPE official blog (10 of the top 15 built by HPE; direct fetch timed out, verified via search snippet of the official page): https://www.hpe.com/us/en/newsroom/blog-post/2026/11/hpe-stacks-10-of-the-top-15-most-powerful-supercomputers-in-the-world.html , 2026-11-16
- StorageReview report (node-hour conversion, Discovery plan): https://www.storagereview.com/news/hpe-lux-ai-cluster-top500-no-14-220-79-petaflops-ornl-genesis-mission , 2026-11-16
- AMD Instinct MI355X official spec page (direct fetch timed out, verified via search snippet of the official page): https://www.amd.com/en/products/accelerators/instinct/mi350/mi355x.html , launched 2025-06-12
- Dataconomy (GPU/node magnitude, agreement value, FP64/FP8/FP4 system peaks; 403, verified via search snippet and cross-checked against calculated values from AMD's official per-card specs): https://dataconomy.com/2026-11-17/amd-powers-40-of-top500-as-lux-debuts-at-no-14/ , 2026-11-17
