Exascale Supercomputers
Frontier passed one quintillion operations a second in May 2022; by June 2026 five systems are exascale, led by China's LineShine at 2.2 exaflops.
Open in the interactive tree →Exascale means at least 10^18 floating-point operations per second at 64-bit precision. Frontier at Oak Ridge National Laboratory reached 1.1 exaflops in May 2022 with more than 9,400 CPU nodes and over 37,000 GPUs, and also led the Green500 list at 62.68 gigaflops per watt. Such machines run climate, materials, nuclear-stockpile and medical simulations and increasingly AI training.
As of October 2026
The June 2026 TOP500 (67th edition) lists five exascale systems: LineShine at the National Supercomputing Centre in Shenzhen (2.198 exaflops, the new No. 1), El Capitan (1.809), Frontier (1.353), Aurora (1.012) and Europe's first, JUPITER Booster (1.000). El Capitan had topped the list since November 2024, when it reached 1.742 exaflops. All 500 systems together now exceed 18.7 exaflops, and El Capitan draws about 30 MW of electricity (TOP500), the new No. 1 LineShine about 42 MW.
Open steps
- Exact arithmetic on AI-built chips Medium AI leverageAI-oriented GPUs favour low precision. Emulating 64-bit matrix maths on them was up to 2.3x faster on GB200 in a test; whether real codes keep their answers is open.
- Power per result at exascale Low AI leverageFrontier led the Green500 at 62.7 GFlops/W in 2022; the June 2026 leader reaches 73.3, while the top ten draw about 205 MW. How can results per watt grow faster?
- AI agents port old simulation codes High AI leverageAgents ported a 74,000-line ocean model from Fortran to GPU-ready C++/Kokkos with staged checks. Can this scale to codes that must match results bit for bit?
- Km-scale climate and AI emulators Medium AI leverageA 3.25 km global atmosphere model ran over a year per day on Frontier; an AI emulator runs about 1,500 years per day at 1 degree. Does it stay right in unseen climates?
Where AI could help
Medium AI leverage. AI training now runs on the same GPU machines, and surrogate models can replace costly simulation steps; power and code porting limit scale.
- Surrogate models that replace the costliest steps of climate and materials simulations
- AI tools that port and tune simulation code for GPUs
- Anomaly detection for failing nodes in machines that draw tens of megawatts
- Active learning that picks which simulations to run next
Shown so far
- Oak Ridge researchers trained language models of up to 1 trillion parameters on 3,072 of Frontier's MI250X GPUs with 100% weak-scaling efficiency (arXiv, December 2023). source
Prerequisites
- Microprocessor1971
- Open Source & Linux1991Linux runs all of the world's top supercomputers
- Graphics Processor (GPU)1999