TOP500 November 2026 (SC26): LineShine, El Capitan, JUPITER and What the New Supercomputer Rankings Mean for AI Infrastructure

TOP500 November 2026 (SC26): LineShine, El Capitan, JUPITER and What the New Supercomputer Rankings Mean for AI Infrastructure

TOP500 November 2026 (SC26): LineShine, El Capitan, JUPITER and What the New Supercomputer Rankings Mean for AI Infrastructure

A CPU-only machine with no GPUs at all is the fastest supercomputer on Earth, while the largest AI clusters ever built do not appear on the list. That paradox frames the TOP500 November 2026 list, the 68th edition, due during SC26 in Chicago this week. In June, China’s LineShine took No. 1 at 2.198 exaflop/s on HPL, ending El Capitan’s run. It did so on custom Arm processors, not accelerators. At the same time, hyperscalers are building GPU campuses measured in gigawatts that never run HPL publicly. For anyone planning AI infrastructure, the rankings are still useful, but only if you read three benchmarks together and know what each one measures.

What this covers: the verified June 2026 baseline, what to watch at SC26 (labelled as expectations), how HPL, HPCG, HPL-MxP and Green500 differ mechanically, LineShine’s architecture with worked FLOPS math, why AI clusters stay off the list, and a planning framework for enterprise AI and HPC buyers.

Status note (16 November 2026, 08:00 IST): At the time of writing, top500.org still shows the June 2026 list as current, and the November 2026 list had not yet been posted. Everything below about the 68th list is framed as what to watch, and every number given is from the verified June 2026 (67th) list unless stated otherwise. We will update this post when the SC26 results are public.

Context and Background

The TOP500 project has ranked the world’s 500 fastest publicly disclosed supercomputers twice a year since 1993. The June list launches at ISC High Performance in Hamburg; the November list launches at SC, the ACM/IEEE supercomputing conference in the United States. SC26 runs 15 to 20 November 2026 at McCormick Place in Chicago, with the exhibition from 17 to 19 November. The list itself is ranked by HPL (High-Performance Linpack), a dense FP64 linear-equation solver.

The TOP500 November 2026 edition is institutionally unusual. At ISC 2026, the Meuer family and ISC Group announced that stewardship of TOP500 passes to ACM, with a new steering committee that includes Jack Dongarra, Erich Strohmaier, Horst Simon and SIGHPC chair Christine Harvey, plus Satoshi Matsuoka, Wu Feng and Anders Jensen. The 68th list is the first produced by the new technical committee. The announcement did not say whether submission rules, methodology or format will change, so treat any claims of a “new TOP500 methodology” with caution until the committee publishes them.

The competitive context changed in June. The 67th list put LineShine at the National Supercomputing Centre in Shenzhen at No. 1. It is the first Chinese system to lead since Sunway TaihuLight in 2017. It also made five systems exceed one exaflop/s on HPL, and for the first time, exascale systems were listed in Asia, North America and Europe simultaneously. JUPITER Booster at Jülich, Europe’s first exascale entry, had crossed exactly 1.000 exaflop/s on the November 2025 list.

Meanwhile, the network and power stories that dominate AI data centres are visible in the top ten. Slingshot-11, InfiniBand NDR and proprietary fabrics such as Tofu D and LingQi all appear. If you want the transport-level background, our comparison of NVLink vs InfiniBand vs RoCE for NCCL all-reduce collectives covers how those fabrics behave under training traffic. This post focuses on what the rankings themselves can and cannot tell you.

The Verified Baseline Before the TOP500 November 2026 Release

The current TOP500 (June 2026) ranks LineShine first at 2.198 EF/s HPL, ahead of El Capitan (1.809 EF/s), Frontier (1.353), Aurora (1.012) and JUPITER Booster (1.000). LineShine also leads HPCG at 22.00 PF/s, but El Capitan leads HPL-MxP at 16.7 EF/s, and France’s KAIROS leads Green500 at 73.28 GFlops/W.

Those four sentences contain the whole thesis of this post: a different system wins each benchmark. The top ten from the official June 2026 TOP500 table is reproduced below with two derived columns. HPL efficiency is Rmax divided by Rpeak; energy efficiency is Rmax divided by reported power. The derived columns are our arithmetic on the published figures, not TOP500 statements.

Rank System Site Rmax (PF/s) Rpeak (PF/s) Power (MW) Rmax/Rpeak GFlops/W (derived)
1 LineShine NSCS Shenzhen, China 2,198.40 2,735.82 42.22 80.4% 52.1
2 El Capitan LLNL, USA 1,809.00 2,821.10 29.69 64.1% 60.9
3 Frontier ORNL, USA 1,353.00 2,055.72 24.61 65.8% 55.0
4 Aurora Argonne, USA 1,012.00 1,980.01 38.70 51.1% 26.2
5 JUPITER Booster Jülich, Germany 1,000.00 1,226.28 15.79 81.5% 63.3
6 HPC7 Eni, Italy 571.50 861.13 8.74 66.4% 65.4
7 Eagle Microsoft Azure, USA 561.20 846.84 not reported 66.3% n/a
8 HPC6 Eni, Italy 477.90 606.97 8.46 78.7% 56.5
9 Fugaku RIKEN, Japan 442.01 537.21 29.90 82.3% 14.8
10 Alps CSCS, Switzerland 434.90 574.84 7.12 75.7% 61.0

TOP500 November 2026 benchmark map showing HPL, HPCG, HPL-MxP and Green500 measuring different bottlenecks on the same supercomputer
Figure 1: One machine, four rankings. Each benchmark stresses a different bottleneck, which is why the leaders differ.

Figure 1 maps each benchmark to the hardware limit it exposes. HPL is a dense LU factorisation in FP64, so it saturates floating-point units and rewards high Rpeak. HPCG solves a sparse system with a preconditioned conjugate gradient, so it is bound by memory bandwidth and network latency. HPL-MxP factorises in low precision and then refines to FP64 accuracy, so it rewards tensor or matrix units. Green500 reuses the HPL run but divides by measured power.

A CPU-only machine leads on two of the four lists

LineShine’s HPL lead over El Capitan is about 21.5% (2,198.4 versus 1,809.0 PF/s). More interesting is how it gets there. Its Rpeak (2,735.8 PF/s) is slightly lower than El Capitan’s (2,821.1 PF/s). The lead comes from efficiency: LineShine converts 80.4% of peak into HPL throughput, while El Capitan converts 64.1%.

That pattern is a known signature of CPU and CPU-like architectures. Fugaku, also CPU-only on the A64FX, posts 82.3%. JUPITER Booster, which uses tightly coupled GH200 superchips, also reaches 81.5%. GPU systems with separate host and device memory, or with very high peak FP64 matrix throughput relative to bandwidth, typically land in the 60 to 70 percent band on HPL. HPL efficiency is not a quality grade, though. It mostly tells you how hard the peak number was to reach.

On HPCG, the June 2026 HPCG table lists LineShine at 22.00 PF/s, El Capitan at 17.41, Fugaku at 16.00 and Frontier at 14.05. As a fraction of HPL, LineShine reaches 1.00%, El Capitan 0.96% and Frontier 1.04%. Fugaku reaches 3.62%. That outlier is the important one: Fugaku was designed around memory bandwidth per FLOP, and the HPCG ratio shows it.

The GPU machines win where AI-style arithmetic is allowed

HPL-MxP inverts the order. According to the TOP500 June 2026 announcement, El Capitan leads at 16.7 EF/s, a 9.2x speedup over its HPL result. Aurora follows at 11.6 EF/s and Frontier at 11.4 EF/s. LineShine debuts fourth at 7.92 EF/s, a 3.6x speedup that TOP500 called modest and attributed to its CPU-only design.

The speedup ratio is the number to watch. It approximates how much more arithmetic the machine can do when the inner loop is allowed to use low precision. Our derived ratios are 9.2x for El Capitan, about 11.5x for Aurora and about 8.4x for Frontier. A ratio near 3.6x says the matrix engines in LX2 do not gain as much from reduced precision as GPU tensor cores do, or that the submission did not exploit it fully. Neither cause is disclosed.

Efficiency leadership sits far down the list

The June 2026 Green500 top three are identical BullSequana XH3000 systems built on NVIDIA GH200 with quad-rail InfiniBand NDR200. KAIROS at CALMIP/University of Toulouse leads at 73.282 GFlops/W, ROMEO-2025 follows at 70.912, and DKRZ’s Levante GPU extension posts 69.426. Isambard-AI phase 1 (68.835) and Paderborn’s Otus GPU partition (68.177) complete the top five.

KAIROS ranks No. 446 on TOP500 and delivers 3.046 PF/s on HPL. Green500’s own note explains why small systems win: computational efficiency tends to decrease slightly with size. Among the top ten TOP500 systems, the best derived efficiency is HPC7 at about 65.4 GFlops/W, and JUPITER Booster reaches about 63.3. LineShine’s 52.07 GFlops/W is respectable for a CPU-only design but well behind accelerated systems.

Deeper Analysis: Inside LineShine and the Arithmetic of the Rankings

The most detailed public description of LineShine comes from the Chips and Cheese ISC26 write-up, which reports details presented at the TOP500 session. We use those figures here and label derived numbers. The Shenzhen Cloud Computing Centre has not disclosed who designed the LX2 processor. Some analysts attribute it to Huawei, but that attribution is unconfirmed.

LineShine fastest supercomputer 2026 hierarchy from LX2 core to 90-cabinet system
Figure 2: The LineShine build-up as reported at ISC26, from Arm core to the full 90-cabinet system.

Figure 2 traces the hierarchy. Each LX2 package has two compute dies; each die has four 40-core clusters with two cores disabled, giving 38 active cores per cluster, 152 per die and 304 per package. Each core is Armv9 with SVE2 (Scalable Vector Extension 2) and SME (Scalable Matrix Extension). A node holds two LX2 packages and connects at 800 Gb/s per CPU, or 1.6 Tb/s per node. Eight nodes form a blade, 16 blades a frame, two frames a cabinet, and 90 cabinets the system.

Worked math: does the reported spec reproduce Rpeak?

A useful habit with any new list entry is to check whether the published component specs reproduce the published Rpeak. The reported per-chip figure is 60.3 TFLOP/s FP64 at 1.55 GHz across 304 cores. Dividing gives:

  • 60.3 × 10^12 / (304 cores × 1.55 × 10^9 Hz) ≈ 128 FP64 FLOP per cycle per core.

That is far beyond what a conventional SIMD pipeline delivers. For comparison, a core with two 512-bit FMA pipes reaches 32 FP64 FLOP per cycle. Reaching 128 implies the SME matrix unit carries most of the FP64 throughput, in the way GPU matrix cores do. This is derived, not disclosed, but it explains why a “CPU-only” machine behaves on HPL more like an accelerator.

Now scale up. The official core count is 13,789,440. Dividing by 608 cores per node gives exactly 22,680 nodes, or 45,360 LX2 packages. Multiplying 45,360 by 60.3 TFLOP/s gives about 2,735 PF/s, which matches the official Rpeak of 2,735.82 PF/s to within rounding. The full mechanical hierarchy of 90 × 2 × 16 × 8 would hold 23,040 node slots, so roughly 360 slots appear unpopulated or not included in the run. Press reports of “about 20,480 nodes” and “over 22,000 nodes” both circulate; the core-count arithmetic supports the higher figure.

Power is the next check. The reported 690 W per LX2 multiplied by 45,360 packages gives about 31.3 MW for processors alone. The measured HPL power is 42.22 MW. That leaves roughly 11 MW, or about 26%, for DDR5, network, storage-adjacent hardware, power conversion and cooling within the measurement boundary. These are illustrative derived figures; the exact Green500 measurement boundary for LineShine has not been published in detail.

Memory per FLOP: the HPCG story in one ratio

Each LX2 has eight stacks of on-package memory, 32 GB total at a reported 4 TB/s, plus 256 GB of DDR5. The bytes-per-FLOP ratio of the fast tier is about 4 TB/s ÷ 60.3 TF/s ≈ 0.066 bytes per FLOP. For a GPU reference, AMD’s MI300A pairs 128 GB of HBM3 at up to 5.3 TB/s with a listed FP64 matrix peak of 61.3 TF/s, which is about 0.086 bytes per FLOP on that published figure.

Those ratios are in the same range, which is consistent with LineShine and El Capitan landing near 1% of HPL on HPCG. Fugaku’s 3.6% HPCG-to-HPL ratio reflects a much higher bandwidth-to-compute balance. The lesson for buyers is direct: sparse solvers, graph analytics and many digital-twin simulations scale with bandwidth, not peak FLOPS. HPCG is the better proxy for those codes.

Why HPL-MxP rewards tensor units

HPL-MxP, formerly HPL-AI, is specified at hpl-mxp.org and described in a 2025 arXiv paper by its maintainers. It performs the LU factorisation in low precision, typically 16-bit, then runs GMRES iterative refinement, using the low-precision L and U factors as a preconditioner, until the solution reaches FP64-level accuracy. The reported rate credits the machine with the FLOPs of an FP64 solve of the same size. A machine that runs FP16 or BF16 matrix multiplies ten times faster than FP64 can therefore post roughly ten times its HPL result, minus refinement overhead.

This is why the speedup ratio maps to hardware design. GPUs devote most of their silicon to low-precision tensor throughput. The Register reported in November 2025 that each MI300A and each GH200 offers roughly 2 PF/s of dense FP8, putting El Capitan’s theoretical FP8 peak near 90 EF/s. Those are theoretical peaks, not measurements. Neither HPL nor HPL-MxP uses FP8 for the official result, so even HPL-MxP understates what an AI training workload sees.

Methodology caveats

Three cautions apply to any table built from these lists, including the TOP500 November 2026 tables when they arrive. First, Rmax figures come from a single optimised run per system, often tuned for weeks; they are not sustained application throughput. Second, the HPL-MxP list has far fewer submissions than TOP500, so ranks are not comparable across lists. Third, power figures depend on the measurement level and boundary each site uses, which Green500 rules allow to vary. Our derived GFlops/W column for TOP500 systems should be read as indicative, not as a substitute for the official Green500 rank.

What to Watch at SC26: Expectations for the TOP500 November 2026 List

Everything in this section is an expectation, not a result, until the TOP500 November 2026 tables are posted. We list the signals worth checking when the TOP500 November 2026 list appears, and why each matters.

Will LineShine hold No. 1? No publicly confirmed system is known to exceed 2.198 EF/s on HPL in time for November. El Capitan’s operators could resubmit an improved run, as sites often do. Closing a 21.5% gap through tuning alone would be unusual. Expect LineShine to remain on top unless a surprise system appears.

Does China submit more systems? Chips and Cheese raised the question of whether Sunway OceanLight and other Chinese national systems will be submitted. The Next Platform noted in November 2025 that some systems, including clusters at Chinese national labs and NCSA, do not submit. Any new Chinese entries would materially shift national share statistics.

Do European systems move? JUPITER Booster posted exactly 1,000.00 PF/s on both the November 2025 and June 2026 lists. JUPITER’s CPU-based Universal Cluster was still under development as of November 2025, and is expected to add only about five petaflops, which would not change the Booster’s rank. France’s Alice Recoque, contracted on 18 November 2025 at €354.8 million with AMD Venice CPUs and Instinct MI430X GPUs, is scheduled to begin installation in 2026. A partial early run is possible but not announced.

Do Vera Rubin systems appear? The US Department of Energy said NERSC’s Dell-built Doudna system, using NVIDIA’s Vera Rubin platform, is due in 2026. Whether it is benchmarked in time for SC26 is not confirmed. If it appears, its FP64 and HPL-MxP figures will be the first public look at Rubin on these benchmarks.

Is the first ACM list different? Watch for changes in how the list presents HPL-MxP, energy and submission metadata. None have been announced.

Why AI Clusters Do Not Appear on the TOP500 November 2026 List

The largest compute installations of 2026 are not on TOP500. Data Center Knowledge reports xAI’s Colossus 2 targets about 2 GW and roughly 555,000 NVIDIA GPUs, and that Microsoft’s Fairwater site in Wisconsin went live in June 2026 at more than 350 MW. Microsoft also claimed its Wisconsin campus hosts the “world’s most powerful supercomputer”, a claim based on different metrics, not on HPL. These systems would dwarf LineShine on AI throughput, and none has a published HPL result.

SC26 supercomputing decision flow showing why hyperscale AI clusters are rarely submitted to the TOP500
Figure 3: The submission decision for a large AI cluster. Most paths end off the list.

Figure 3 shows the decision chain. It starts with intent: a public ranking has little commercial value for a private training cluster, and disclosure can reveal capacity to competitors. Even with intent, the hardware must have strong FP64. Then the owner must take the whole cluster offline for many hours of HPL, and the system must run stably at full scale. Each gate removes candidates.

FP64 is shrinking on AI silicon

The FP64 gate is newer and more decisive than it looks. According to Glenn Lockwood’s compilation of NVIDIA’s datasheets, the B200 offers about 37 TF/s of FP64, while the Blackwell Ultra B300 offers about 1.2 TF/s. The same B300 is listed at roughly 4,500 TF/s of dense FP8. NVIDIA traded FP64 silicon for low-precision throughput because AI training rarely needs FP64.

A worked illustration makes the effect concrete. Treat these as order-of-magnitude numbers from published peaks, not benchmark results:

Illustrative cluster FP64 peak per GPU Cluster FP64 peak Dense FP8 peak
1,024 × B200 ~37 TF/s ~38 PF/s not computed here
1,024 × B300 ~1.2 TF/s ~1.2 PF/s ~4.6 EF/s

A thousand-GPU B300 cluster would offer about 4.6 exaflops of dense FP8 but only about 1.2 petaflops of FP64 peak. The Next Platform notes a machine now needs more than 10 PF/s of HPL to make the list at all. The B300 cluster above could not qualify on FP64, even though it is an extremely capable AI system. As more AI fleets move to FP64-light parts, the overlap between “big AI cluster” and “TOP500 system” will shrink further. HPL-MxP partially bridges that gap, but it still needs a usable FP64 path for refinement.

The cost of idling a cluster

HPL at exascale runs for hours, and tuning runs add more. For a hyperscaler, that time is lost revenue or lost training progress. As an illustrative calculation, suppose a 10,000-GPU cluster is billed internally at $2 per GPU-hour. A 24-hour benchmark window then costs about $480,000 in forgone capacity, before engineering time. That figure is a hypothetical, not a reported price, but the order of magnitude explains why only a few cloud systems, such as Microsoft’s Eagle at No. 7, appear.

What the list still tells AI buyers

The absence does not make TOP500 irrelevant for AI teams. The list remains the only large, audited, comparable record of how real systems behave at scale, including network topology, power and cooling. For the network layer, the top ten split between HPE Slingshot-11, InfiniBand NDR, and proprietary fabrics. Our deep dive on RDMA, RoCE v2, InfiniBand and Ultra Ethernet congestion control explains why those choices matter more for collective-heavy training than for HPL.

Power: The Constraint Every List Now Reflects

Power is the most transferable lesson from the TOP500 November 2026 cycle. LineShine draws 42.22 MW for 2.198 EF/s. El Capitan draws 29.69 MW for 1.809 EF/s. Aurora draws 38.70 MW for 1.012 EF/s. The spread is close to 2.5x in FLOPS per watt among exascale machines that entered service within a few years of each other.

Here is a planning-scale illustration. A facility with a 30 MW IT envelope could host about 1.83 EF/s of HPL at El Capitan’s derived 60.9 GFlops/W, about 1.56 EF/s at LineShine’s 52.07, or about 0.79 EF/s at Aurora’s 26.2. The building and the grid connection are fixed costs, so efficiency directly sets the useful compute per site. This is illustrative arithmetic on published numbers, not a vendor quote.

At hyperscale, the same logic pushes into power architecture. Gigawatt AI campuses are moving to higher-voltage DC distribution to cut conversion losses. Our analysis of the 800 V DC AI data-centre power architecture and solid-state transformers covers that transition. The Green500 measures efficiency at the system level, but facility-level losses from transformers to chips decide what you pay for each useful FLOP.

Cooling follows the same curve. Every system in the Green500 top five uses direct liquid cooling or a platform designed for it, and Alice Recoque is specified with warm-water direct liquid cooling for its unified racks. Air-cooled designs are no longer competitive at the top of either list.

Trade-offs, Gotchas, and What Goes Wrong

Reading HPL as application speed. The headline of the TOP500 November 2026 list will be an HPL number. HPL measures dense FP64 linear algebra under ideal tuning. Most production codes run at a small fraction of HPL. The HPCG ratios above, around 1% of HPL on the top GPU and LX2 systems, are closer to what memory-bound simulation codes see. If your workload is a sparse finite-element model or a large digital twin, HPL rank is a weak predictor.

Treating HPL-MxP as an AI benchmark. HPL-MxP is still a linear-equation solver. It shows low-precision matrix throughput, but not attention kernels, all-to-all expert routing, checkpoint I/O or data-loader stalls. MLPerf Training and Inference cover those behaviours better. HPL-MxP’s value is that it runs on scientific systems that never submit to MLPerf.

Comparing Rmax across years without context. Systems are resubmitted after tuning or expansion, so a changed Rmax may not mean new hardware. JUPITER posted the same 1,000.00 PF/s on two consecutive lists, which suggests no resubmission, though the list does not say so.

Ignoring the measurement boundary in power. Green500 rules allow different measurement levels and boundaries. Two systems with the same GFlops/W may have included different amounts of the facility. Use derived TOP500 efficiency only as a rough signal.

Assuming the list is complete. Unsubmitted Chinese systems, unsubmitted national systems elsewhere and nearly all hyperscale AI clusters are absent. National share statistics are therefore statistics about submitted systems, not about installed compute.

Over-reading LineShine’s vendor story. The LX2’s designer has not been disclosed, and claims about its memory technology are speculation. Chips and Cheese explicitly labels its view that the HBM is likely not conventional HBM as an opinion.

Green500 November 2026 and HPL-MxP AI benchmark planning flow for choosing which ranking to trust
Figure 4: A planning flow for deciding which ranking to weight, based on precision and memory needs.

Practical Recommendations

Use the TOP500 November 2026 rankings as calibrated evidence, not as purchasing advice. Figure 4 summarises the approach: start from your workload’s precision needs, decide whether it is memory-bound, and only then pick which benchmark to trust. A climate model, a CFD-heavy product digital twin and an LLM fine-tuning pipeline need three different readings of the same lists.

For FP64 simulation, weight HPCG and FP64 efficiency over HPL rank. Ask vendors for HPCG results on configurations close to yours, and for sustained memory bandwidth per FP64 FLOP. If you are buying AI-first GPUs, check the FP64 datasheet line before assuming they can run your simulation codes; the B200 to B300 change shows that line can drop by an order of magnitude.

For AI training and inference, treat HPL-MxP as a hint and MLPerf plus your own pilot as the evidence. Network design, scheduler integration and checkpoint storage will dominate realised throughput. Our Slurm vs Kubernetes guide for AI training covers the scheduling side.

For facilities, plan in megawatts first. The spread from about 26 to about 65 GFlops/W in the top ten means efficiency, not FLOPS, sets the ceiling of a fixed-power building.

Checklist:

  • Classify workloads by precision (FP64, mixed, FP8/FP4) and memory-boundness.
  • For FP64 work, compare HPCG-to-HPL ratios and bytes per FLOP.
  • For AI work, verify per-GPU FP64 if any simulation must share the fleet.
  • Model capacity per megawatt, not per rack.
  • Specify liquid cooling readiness for any new accelerated deployment.
  • Pilot your own top three applications before committing to a vendor.
  • Re-check the TOP500 November 2026 tables when published, especially new entries and HPL-MxP ratios.

Frequently Asked Questions

When is the TOP500 November 2026 list released?

The 68th TOP500 list is scheduled for release during SC26, which runs 15 to 20 November 2026 in Chicago. Historically the November list is announced early in SC week, around the opening of the exhibition. As of the morning of 16 November 2026, top500.org still shows the June 2026 list as current. This is also the first list produced under ACM stewardship, following the handover announced at ISC 2026. Check top500.org directly for the official tables rather than relying on early press summaries.

What is the fastest supercomputer in 2026?

On the verified June 2026 TOP500 list, the fastest supercomputer is LineShine at the National Supercomputing Centre in Shenzhen, China, with 2.198 exaflop/s on HPL. It uses custom 304-core Armv9 LX2 processors at 1.55 GHz, no GPUs, a proprietary LingQi interconnect and Kylin OS. El Capitan at Lawrence Livermore is second at 1.809 exaflop/s. On mixed-precision HPL-MxP, however, El Capitan is fastest at 16.7 exaflop/s. “Fastest” therefore depends on which benchmark you use.

What is the difference between HPL, HPCG and HPL-MxP?

HPL solves a dense system of linear equations in FP64, which rewards raw floating-point throughput and ranks the TOP500. HPCG solves a sparse system with a preconditioned conjugate gradient method, which stresses memory bandwidth and network latency, so results are usually around 1 to 4 percent of HPL. HPL-MxP factorises the matrix in low precision such as FP16, then refines the answer to FP64 accuracy using GMRES. That rewards tensor and matrix units, so GPU systems often post eight to twelve times their HPL score.

Why is a CPU-only system faster than GPU supercomputers on HPL?

LineShine has slightly lower theoretical peak than El Capitan but converts more of it into HPL throughput: 80.4% versus 64.1%. Derived arithmetic suggests each LX2 core delivers about 128 FP64 FLOP per cycle, which implies a matrix unit similar in role to a GPU’s matrix cores. So it is “CPU-only” in programming model but matrix-accelerated in practice. On low-precision HPL-MxP, its advantage disappears, and it ranks fourth with a 3.6x speedup versus El Capitan’s 9.2x.

Why are AI supercomputers like Colossus not on the TOP500?

Submitting is voluntary, and hyperscale owners gain little from a public HPL score. A full HPL run idles the whole cluster for hours, which costs training time or revenue. Newer AI GPUs also have reduced FP64: NVIDIA’s B300 lists about 1.2 TF/s FP64 versus about 37 TF/s on B200, so HPL results would understate their AI capability. Disclosure can also reveal capacity to competitors. A few cloud systems, such as Microsoft’s Eagle, do submit, but they are exceptions.

Which system leads the Green500 in 2026?

On the June 2026 Green500, KAIROS at CALMIP and the University of Toulouse leads with 73.282 GFlops/W. It is a BullSequana XH3000 system with NVIDIA GH200 superchips and quad-rail InfiniBand NDR200, ranked No. 446 on TOP500. ROMEO-2025 (70.912) and DKRZ’s Levante GPU extension (69.426) use the same architecture. Smaller systems tend to lead because efficiency drops slightly with scale. The Green500 November 2026 list will be published alongside the TOP500 at SC26.

Further Reading

By Riju — about

Comments

No comments yet. Why don’t you start the discussion?

Leave a Reply

Your email address will not be published. Required fields are marked *