
Top500.org
The 68th Top500 list, published Monday at the SC26 conference in Chicago, shows the average listed supercomputer now has more than 445,000 cores, up about 46 percent from 305,404 in June, while the ten fastest machines did not change places. Readers comparing systems should not read that jump as a speed gain: combined performance on the list rose only about 5 percent over the same six months, and Next Platform reports that Top500 core totals count streaming multiprocessors on GPUs as cores.
Only 27 new systems joined the list, the fewest in its history, and the entry threshold rose 7.3 percent from June to 2.857 petaflops, according to Network World.
China's LineShine, installed at the National Supercomputing Center in Shenzhen, still leads at 2.198 exaflops on the High Performance Linpack (HPL) benchmark. El Capitan, Frontier, Aurora and Europe's JUPITER Booster follow, and they remain the only five systems to reach one exaflop.
The list's core column adds CPU cores and accelerator cores into one total per system. Next Platform, which analyzes the new machines on each list, says that total counts streaming multiprocessors on GPUs alongside the cores in CPUs.
Frontier shows how lopsided that tally gets. Of its 9,066,176 listed cores, only about 600,000 are AMD Epyc CPU cores, and more than 8 million come from Instinct MI250X GPUs, per ComputerBase. By TechTimes arithmetic, GPU hardware makes up roughly 93 percent of that total.
LineShine's count is a different unit. It uses 45,360 LingKun LX2 processors, each with 304 Arm v9.2 cores, which multiplies out to 13,789,440 cores that are all CPU cores. A core total therefore mixes CPU cores and GPU compute blocks, and it is not an input to the ranking. The Top500 ranks systems by HPL, which solves dense systems of linear equations in 64-bit arithmetic.
Read more: China Seizes TOP500 Crown With All-Domestic Chips: LineShine Tops List, Trails US on AI
Top500's own highlights page puts the average at 445,578 cores per system, up from 305,404 six months ago. ComputerBase counts 445,441. Multiplied across 500 systems, that implies roughly 223 million listed cores against roughly 153 million in June, or about 70 million more. That figure is TechTimes arithmetic, not a number the Top500 publishes.
Speed barely moved by comparison. Network World reports combined HPL performance up 4.4 percent to 19.57 exaflops. The Top500 highlights page lists 19.70 exaflops against 18.73 six months ago, about 5.2 percent by TechTimes arithmetic. The two sources differ slightly, but both show growth far below the core-count jump.
ComputerBase also reports that 293 hybrid systems, or 58.6 percent, now pair CPUs with accelerators, up from 276 (55.2 percent) in June. Top500's highlights count 294 accelerated systems, including 110 with Nvidia Hopper, 61 with Nvidia Ampere and 34 with AMD Instinct, and Network World counts 291. A share shift of roughly 3.4 to 3.6 percentage points, depending on the source, looks small beside a 46 percent jump, and the sources read for this article do not break down where the added cores sit.
Read more: South Korea Launches Hangang Supercomputer: Quantum Computing Integration Due by May
LineShine reaches about 80 percent of its 2.736-exaflop theoretical peak and runs on a proprietary interconnect called LingQi.
The lead carries limits. LineShine draws 42.2 megawatts and delivers 52.07 gigaflops per watt, against El Capitan's 60.94. On HPL-MxP, the mixed-precision benchmark used as an AI proxy, it ranked fourth in June at 7.92 exaflops, behind El Capitan at 16.7, Aurora at 11.6 and Frontier at 11.4. LineShine is installed at a national supercomputing center, not a commercial service, and China's 2017 National Intelligence Law requires organizations and citizens to support state intelligence work in accordance with the law.
Only two top-ten machines use no accelerators: LineShine and Japan's Fugaku, which has spent about 6.5 years in the top ten.
The Green500 ranks systems by HPL performance per watt, and its top ten has not changed since June. KAIROS in France leads at 73.282 gigaflops per watt, followed by France's ROMEO-2025 at 70.912 and Germany's Levante GPU extension at 69.426. The leaders are small by Top500 standards: KAIROS delivers 3.05 petaflops on HPL, roughly one seven-hundredth of LineShine's output by TechTimes arithmetic.
ComputerBase's tally shows the United States with 162 systems, unchanged from June, Germany with 47 (up six), Japan with 46 (up two), South Korea with 22 (up three) and France with 20 (down one). JUPITER Booster remains Europe's fastest system at exactly one exaflop.
TechTimes reported on Nov. 2 that South Korea's science ministry expected its new Hangang supercomputer to place in or near the top ten. With the top ten unchanged, Hangang did not break into it.
The list is built from site and vendor submissions only, and Wikipedia notes that Meta's disclosed H100 datacenters have not been benchmarked for it. The Register estimates that Colossus, the xAI cluster in Memphis with about 200,000 Hopper GPUs and a 300-megawatt draw, would pass 3 exaflops on HPL at 25 percent efficiency if its operator submitted it. That estimate sits at the top of the per-GPU range The Register cites, 34 to 67 teraflops of 64-bit performance. At the low end, the same inputs give about 1.7 exaflops, by TechTimes arithmetic.
Top500 organizers say growth "has indeed shifted toward AI factories, and from capability systems to capacity systems." They cite Epoch AI's estimate that five US hyperscalers owned 71 percent of global AI compute capacity as of the fourth quarter of 2025.
Oak Ridge National Laboratory's new Lux system, built with HPE and AMD, delivers 220 petaflops on HPL, and The Register expects it to rank 14th on the list.
The Department of Energy is on track to deploy at least nine new supercomputers under its Genesis Mission, The Register reports. Discovery at Oak Ridge is an HPE Cray machine with AMD Venice CPUs and Instinct MI430X GPUs that The Register says will deliver three to five times Frontier's performance. ORNL's October 2025 announcement put its delivery in 2028.
That timeline puts Discovery outside the June 2027 list, so any change at No. 1 would more likely come from a large AI cluster whose operator chooses to run HPL. For readers weighing the numbers, the November list says three things: core counts add GPU compute blocks and CPU cores into one total, so the 46 percent jump is not a 46 percent speed gain; the ten fastest systems and the Green500 leaders did not move; and the list reflects only the systems whose owners submit them. LineShine's 2.198 exaflops is the top of that submitted set as listed by the Top500, and its 42.2-megawatt draw, fourth-place AI-benchmark result and lower efficiency than El Capitan are part of the same record.
Each system's total adds CPU cores and accelerator cores, and Next Platform reports that GPU streaming multiprocessors are included. That means a GPU-heavy system and a CPU-only system report different kinds of cores under the same column.
By Rmax, the best 64-bit performance each system reaches on the HPL benchmark, with ties broken by theoretical peak. Wikipedia notes critics say HPL tests only dense linear systems, which are not representative of all scientific computing.
El Capitan, the US system at Lawrence Livermore National Laboratory, ranks second at 1.809 exaflops and is the fastest American machine on the list.
Top500 publishes a new list every June and November, so the next edition is due in June 2027.
