China’s 1.5 Exaflops Oceanlite Supercomputer Chases the Gordon Bell Prize

Background

On September 16, the Association for Computing Machinery (ACM) released the shortlist for the Gordon Bell Prize, often called the "Nobel Prize of Supercomputing". A paper describing a 1.5 exaflops Chinese supercomputer, Oceanlite (the upgraded TaihuLight), has been selected and the award ceremony will take place at SC23 in Denver in November 2023.

The nominated paper, titled "Precise Calculation of Turbine‑Machinery Flow", was authored by Weiqi Shen, Jiahuan Cui, Yao Zheng and 16 others from Zhejiang University, Tsinghua University, the National Supercomputing Center in Wuxi, Cambridge University and the University of Florida. The team developed a new large‑eddy‑simulation code for compressible flow in turbine machinery and applied it to a major NASA challenge, solving 1.69 billion mesh elements and 865 billion degrees of freedom with a high‑order unstructured solver.

Oceanlite is built around more than 100 000 custom SW26010 Pro processors designed by the National Research Center of Parallel Computer Engineering (NRCPC). The SW26010 Pro is an improved version of the original SW26010, featuring six core modules per chip (each with one management core and 64 compute cores) and a 256 KB L2 cache per 8×8 compute‑core group. Each core module includes a DDR4 controller and 16 GB memory, giving the chip 96 GB total memory and 307.2 GB/s bandwidth. The processor delivers 14.03 PB of FP64/FP32 precision and 55.3 PB of BF16/FP16 precision.

Oceanlite’s largest configuration comprises 107 520 nodes (41 930 000 cores) across 105 cabinets, with a theoretical peak performance of 1.5 exaflops. Scaling to 120 cabinets would reach 1.72 exaflops (FP64), surpassing the U.S. Frontier system, while 160 cabinets could approach 2.3 exaflops, potentially beating Aurora and El Capitan.

Other Supercomputer Nominees

1. Hybrid ML‑HPC approach – A team led by Sambit Das and others (University of Michigan, Indian Institute of Science, Oak Ridge) combined density‑functional theory with quantum many‑body methods, achieving high‑precision quantum simulations that utilized up to 60 % of the Frontier system.

2. Multi‑physics nuclear‑reactor simulation – Researchers from Penn State, ORNL, ANL and UIUC simulated an advanced reactor coupling radiation transport, heat, and fluid dynamics using the Shift Monte‑Carlo code and NekRS CFD solver on Frontier, solving 10 billion spectral elements and 350 billion degrees of freedom.

3. Turbulent Rayleigh‑Bénard convection – A consortium including KTH, Friedrich‑Alexander‑University, Max‑Planck‑Computing‑and‑Data‑Facility and Ilmenau University used the GPU‑optimized Neko code on LUMI (Finland) and Leonardo (Italy) with up to 16 384 GPUs for unprecedented direct numerical simulations.

4. Seismic processing on Cerebras CS‑2 – Hatem Ltaief, Yuxi Hong and colleagues leveraged the AI‑optimized Cerebras CS‑2 system’s massive memory bandwidth to perform large‑scale seismic data processing via low‑rank matrix approximations and multi‑dimensional convolutional wave‑equation solvers.

5. Deep equivariant models for biomolecular simulation – A Harvard team (Albert Musaelian et al.) introduced the Allegro architecture, achieving quantum‑level accuracy for protein dynamics on the Perlmutter supercomputer, scaling to 100 million atoms and nanosecond‑scale simulations of HIV capsid.

The ACM also announced its first Gordon Bell Climate Modeling Prize. The SCREAM variant of the Energy Exascale Earth System Model, developed at Sandia National Labs, was nominated and runs on Frontier, simulating 1.26 years of climate per day.

China’s Oceanlite system was also a finalist, modeling the 2021‑2022 Tonga submarine volcanic eruption with 400 billion particles and 80 % computational efficiency on 39 million cores.

Jack Dongarra, Turing‑Award‑winning professor at the University of Tennessee, noted that China continues to produce the most supercomputers worldwide, despite limited visibility due to U.S. sanctions. He warned that China’s leadership in exascale computing could provoke stricter export controls.

Historically, China’s Tianhe‑2 surpassed the U.S. Titan in 2013, prompting U.S. bans on Intel chips for upgrades and later entity‑list restrictions in 2021.