AlphaGazette
Menu

Tech

LineShine stays first on Top500 at 2.2 exaFLOPS, ahead of El Capitan

AMD chips power 40% of the new ranking, with 202 systems, and Oak Ridge’s Lux enters at position 14. A larger Discovery machine is still planned under the Genesis Mission.

2 min read AI-generated

China’s LineShine remains first on the new Top500 list of supercomputers, at 2.2 exaFLOPS on the High Performance Linpack benchmark, a lead of roughly 22 percent over El Capitan at Livermore National Laboratory. The Shenzhen machine holds that place on Chinese processors after US export controls blocked American GPUs.

The system at the top

The National Supercomputing Centre in Shenzhen built LineShine from 45,360 LingKun LX2 CPUs. Each processor has 304 Arm v9.2 cores; one description clocks them at 1.5 GHz and the list’s system line at 1.55 GHz, with LingQi parts and Kylin OS. Drawing 42.2 megawatts, the lead carries a clearly higher power bill. El Capitan, on AMD chips, stays second. Frontier is third, Aurora fourth and JUPITER Booster fifth, then HPC7, Microsoft’s Eagle and HPC6. Those places match the previous ranking.

AMD’s share and Lux

AMD powers 40% of the list, 14% higher than in 2025. AMD-based machines now number 202, against 177 last year. El Capitan, Frontier, HPC7 and HPC6 give AMD four of the top 10 spots it already held.

New at position 14 is Lux, housed in Oak Ridge National Laboratory’s Oak Ridge Leadership Computing Facility beside Frontier. It is the first of two machines from a $1 billion arrangement a little over a year ago between AMD and the Department of Energy, and its first appearance on the Top500. HPE built it with AMD and the department as part of the Genesis Mission. The machine uses over 4,000 Instinct MI355X GPUs in over 500 nodes, AMD EPYC host CPUs and Pensando networking, inside HPE ProLiant Compute XD685 nodes.

On Linpack, Lux delivers 220 petaFLOPS of double-precision work; two years ago that class of result would have placed in the top 10. A separate FP64 figure for HPC workloads is 317 petaFLOPS. Lower-precision numbers do not line up: one puts FP4 peak performance somewhere north of 20 exaFLOPS, while another lists 20 exaFLOPS of FP8 for training and 40 exaFLOPS of FP4 for inference. Each GPU carries 288 GB of HBM3E, and the full system can reach 1.1 PB of that memory.

What comes after this list

The Genesis Mission, announced late last year, has the Department of Energy on course to field at least nine more supercomputers. Most are smaller or aimed at AI, with few truly large systems. Oak Ridge’s coming Discovery machine, an HPE Cray design, is to combine AMD Venice CPUs with Instinct MI430X GPUs that still emphasize double precision rather than low-precision AI throughput. Each of those multi-die chips is estimated north of 200 teraFLOPS for FP64 matrix math. The department says Discovery is to reach between 3x and 5x Frontier’s performance.