ORNL's first Genesis Mission supercomputer misses the top ten even at perfect Linpack efficiency. No US system has published the FP64 to challenge China's LineShine.
Oak Ridge National Laboratory's Lux supercomputer comes online this October. The lab set a date on August 21: the new system, powered by AMD and built by HPE, "combines AI and high-performance computing" in service of the Department of Energy's Genesis Mission, which counts Lux as its first designated machine. The announcement says nothing about the TOP500. No DOE or ORNL statement about the machine has been made.
OLCF publishes throughput three ways: 40 exaflops of FP4 for AI inference, 20 exaflops of FP8 for AI training, and 317 petaflops of FP64 for modeling and simulation.
Only the third figure is eligible for the TOP500. China's LineShine, at the National Supercomputing Center in Shenzhen, took the top spot in June with a measured High Performance Linpack result of 2,198.4 petaflops. Lux's theoretical FP64 peak is 14.4 percent of that measured score, and Linpack only ever subtracts from the theoretical peak.
The distance between Lux's own two headline figures is the more interesting number: 40 exaflops of FP4 against 317 petaflops of FP64, a ratio of 126 to 1. DOE calls both new Oak Ridge machines AI supercomputers, and the hardware is proportioned to match.
OLCF lists Lux at more than 500 nodes and more than 4,000 AMD Instinct MI355X GPUs on an Ultra Ethernet fabric, and quotes 78 teraflops of FP64 per GPU. Eight GPUs to a node puts the configuration at roughly 504 nodes and 4,032 accelerators, which reproduces the published 317 petaflops. That is an inference from a rounded figure, not a count OLCF publishes, and OLCF notes full-system jobs may run on fewer than 4,000 GPUs, with nodes held for persistent services.
Rank comes from Rmax, the measured Linpack score; Rpeak, the theoretical ceiling, only bounds it, and nothing reaches its peak. Which is why the bound is enough here. At a physically impossible 100 percent efficiency, 317 petaflops ranks twelfth on the June 2026 list, below LUMI's 379.70. The top ten is arithmetically out of reach. The list carries two MI355X systems, Sunrise and Zenith at Cambridge, and both measured 75.32 percent; at that rate Lux would post about 239 petaflops and land thirteenth, just behind Leonardo's 241.20.
The MI300X delivers 163.4 teraflops of FP64 matrix and 81.7 of FP64 vector; its CDNA 4 successor delivers 78.6 of each. AMD halved double-precision matrix throughput between generations while leaving vector throughput almost untouched, and published both numbers on its own spec sheets.
The second-place machine is not short of hardware. El Capitan, at Lawrence Livermore, posts an Rmax of 1,809.00 petaflops against an Rpeak of 2,821.10. LineShine's Rpeak is 2,735.82. The US flagship has more theoretical peak than the system that displaced it and lost the efficiency seat.
Matching LineShine's score would take 77.9 percent Linpack efficiency. El Capitan's last two submissions came in at 64.1, unchanged across the 66th and 67th lists. That is not an exotic figure, and it is not a ceiling: JUPITER Booster, a GH200 system on the same list, measured 81.55 percent. What is missing is any announced remeasurement or expansion establishing the gain as forthcoming. The hardware route needs 3.43 exaflops of peak at current efficiency, a 21.5 percent expansion that appears on no public record. None of which is a criticism; El Capitan is an NNSA stockpile simulator, and its list position is a byproduct of that job.
The October 2025 announcement of both Oak Ridge machines promised to "Accelerate American Dominance in Science and Technology." Neither that release nor ORNL names the TOP500 or Linpack, and neither claims a ranking for Lux or Discovery. Both reach for rank anyway. DOE's calls Frontier "the world's second largest supercomputer." ORNL records that Jaguar, Titan, Summit and Frontier "were recognized as the world's fastest system of its time," a claim about lineage, not about either new machine. What the releases foreground for Lux and Discovery is AI capacity, and neither names the metric that produced the heritage.
The August 21 post that fixed the October date runs the same way: AI and high-performance computing for the Genesis Mission, no benchmark named. A targeted search on August 25, 2026 turned up no statement from ORNL, OLCF, AMD, HPE or DOE about a Lux Linpack submission, in either direction. Timing does not obviously preclude one: AMD posted photographs of the fully installed system on August 20, and an October start sits close against TOP500's deadlines, which run to October 24 for new-site submissions with installation by November 1. But eligible is not tuned, run, and submitted, and on those the record is empty.
NNSA confirmed in July that HPE will build Mission and Vision at Los Alamos on NVIDIA's Vera Rubin platform, which "will support both high-performance computing (HPC) and agentic artificial intelligence (AI) workloads," with full deployment "expected in 2027 and 2028." SCN's reporting on Mission traces it to the ATS-5 technical-requirements document, which calls for a 2027 deployment to replace Crossroads. NVIDIA markets Vera Rubin against the TOP500 by name, advertising "5 petaflops of native FP64 support" across as many as 144 GPUs in a rack, and names NERSC's Doudna alongside Mission, Vision and Veritas. Doudna is specified at "more than ten times the performance of Perlmutter," converged for simulation, AI training and inference.
So the pipeline buys native double precision. What no announcement describes is enough of it in one place. Beating 2,198.4 petaflops at a realistic 75 percent Linpack efficiency takes roughly 2,931 petaflops of FP64 peak: about 586 Vera Rubin racks, or 84,000 GPUs, on one fabric tuned for one benchmark. Nothing at that scale has been announced at a US site. NNSA and LANL have published no Linpack target for Mission, and NERSC none for Doudna. An absent target is not evidence a machine cannot win; it is the reason nobody can presently say one will.
Argonne's Equinox and Solstice, at 10,000 and 100,000 NVIDIA Blackwell GPUs, are billed by DOE and NVIDIA at a combined "2,200 exaflops of AI performance", an AI-precision aggregate no primary breaks down by format. Neither has said which Blackwell part is involved. NVIDIA's HGX specifications show why that matters: B300 posts 10 teraflops of FP64 across an eight-GPU platform, compared with 296 for B200, or 1.25 per GPU versus 37. A hundred thousand GPUs is 125 petaflops of FP64 at the first rate and 3,700 at the second. SCN reported in March that Solstice targeted Q2 2026 operational status, and nothing has been published since.
The same page settles a question SCN has been tracking through the FP64 debate. NVIDIA now publishes native and emulated double precision as separate line items for Rubin: 33 teraflops of FP64 per GPU, and 200 teraflops of FP64 DGEMM, footnoted as "peak performance using Tensor Core-based emulation algorithms." NVIDIA is publishing an emulated double-precision figure on its own spec sheet and labeling it as emulated.
That is what makes Discovery the credible shot rather than the only double-precision machine in the pipeline. Announced for 2028 at Oak Ridge, it pairs HPE's Cray GX5000 with AMD EPYC "Venice" processors and Instinct MI430X accelerators, which AMD rates at 288 teraflops of hardware-based FP64 vector, a July 2026 engineering projection. That is about 3.7 times the MI355X, and at that rate the 2,931-petaflop target needs roughly 10,200 accelerators instead of 84,000. DOE has published no figure for Discovery beyond "significantly greater performance than Frontier."
The most detailed public account of the machine is Jack Dongarra's technical report ICL-UT-26-01, dated June 23. It records the Linpack run at roughly 2.198 exaflops, about 80 percent of peak, drawing 40 to 42 megawatts on a matrix of n = 37,647,359. HPCG comes in at 22.0049 petaflops, and HPL-MxP, the list's mixed-precision benchmark, reaches about 7.92 exaflops: one architecture leading three benchmarks that reward three different things.
The LX2 socket underneath is Armv9.2 with 304 cores, SVE2 and SME vector and matrix extensions, 32 GB of on-package HBM at roughly 4 TB/s, and 60.3 teraflops of FP64 in a 690-watt envelope. Dongarra calls it "not simply another GPU-based exascale system," distinguished by "a CPU-centric architecture with built-in matrix engines and high-bandwidth memory, paired with a co-designed software stack."
Two details belong in the record without a verdict. Three authoritative sources give the system three sizes: NSCC-SZ's own arXiv paper puts it at 20,480 nodes, the TOP500 entry records the Linpack run at 22,680, and Dongarra writes "22,000+ nodes." Neither states which machine was used. Dongarra also reports 14 Gordon Bell submissions for SC26 running on LineShine, three of them Prize finalists; ACM has not published its 2026 finalists.
State media has been less circumspect. Xinhua ran commentary in late June under the headline "Tech curbs fail to stop China's supercomputing rise." That is attributable to state messaging. As of August 25, 2026, no public evidence indicates a new or expanded Chinese submission for the 68th list.
Ranking is not inventory. Counting every entry on the June 2026 list, US machines carry 7.03 exaflops of measured Linpack across 161 systems, versus China's 2.38 across 31; submissions have been voluntary and strategic on both sides for years, a point worth repeating from June. Twenty-nine of China's 31 entries sit near the threshold, many anonymized. A US hyperscaler could enter and win at any time; none has submitted since Microsoft's Eagle in 2023.
The deeper problem for the list is that its benchmark is drifting away from the machines being bought. DOE calls both new Oak Ridge systems AI supercomputers, and Lux's 126-to-1 ratio of FP4 to FP64 is that label rendered in silicon. NVIDIA's spec sheets tell the same story, with B300 carrying roughly a thirtieth of B200's double precision per GPU, and AMD's tell it more quietly, with FP64 matrix throughput halved between generations. High Performance Linpack rewards exactly the arithmetic the current procurement cycle is spending least on.
LineShine's own results show how much the answer depends on the question. The machine that measured 2,198 petaflops on Linpack measured 22 on HPCG, the companion benchmark built around the memory-bound access patterns of production simulation codes, and about 7,920 on HPL-MxP in mixed precision. One machine, three benchmarks, a 360-fold spread. And the FP64 line itself now carries an asterisk: NVIDIA publishes native and emulated double precision side by side, which leaves the hardware meaning of a future Linpack score inside the FP64 debate.
Some centers say plainly that ranking is not why they bought. The Maui High Performance Computing Center ribbon-cut Makau this month at 2.64 peak petaflops, below the June list's entry threshold, at a center whose stated mission is evaluating early production technology rather than delivering production capability.
Japan's position is quieter than it is usually rendered. RIKEN's 275-page basic design report for FugakuNEXT never uses the words TOP500 or Linpack, and mentions HPL once, inside a summary of an NVIDIA pitch. Its published targets are set in FP64 vector performance for existing applications and 50 exaflops or more for AI processing. Satoshi Matsuoka told Nikkei xTECH that world-class ranking was not the goal, which SCN reported as a machine that does not chase the list. Nothing in the record says RIKEN will decline to submit.
What the list still owns is continuity: the same measurement, run the same way, in public, twice a year since 1993. Nothing else in the field carries a 33-year longitudinal record under a stable methodology, and that is worth keeping even as what it measures narrows. But the list was never especially good at saying who has more compute, and it is getting worse at it as the machines diversify. What the 68th edition will say cleanly is which countries are still building systems tuned for dense double-precision linear algebra at national scale, and are still willing to be measured doing it. In November, that is China and Europe, plus Fugaku, with the US answer arriving in 2027 and 2028 and no Linpack target published.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
