Sunway TaihuLight

Sunway TaihuLight
Name	Sunway TaihuLight
Country	China
Developer	National Research Center of Parallel Computer Engineering & Technology
Debuted	2016
Peak	93.01 PFLOPS (LINPACK)
Architecture	SW26010 manycore
Processors	40,960 nodes; 260 cores per node
Operating system	Sunway RaiseOS (Linux-like)

Sunway TaihuLight Sunway TaihuLight is a Chinese supercomputer developed for high-performance computing tasks and national initiatives, installed at the National Supercomputing Center, Wuxi, and announced during the 2016 era of international TOP500 list competition. The system played a central role in China's strategic computing programs involving institutions such as the National Research Center of Parallel Computer Engineering & Technology, the Shanghai Supercomputer Center, and collaborations with provincial authorities in Jiangsu. It attracted attention from the United States Department of Energy, the European Commission, and academic groups at Tsinghua University, Peking University, and the Chinese Academy of Sciences for large-scale simulations in climate, materials, and engineering.

Overview

Sunway TaihuLight is a massively parallel system designed to achieve exascale-class workloads through manycore design, built under initiatives similar to projects at Lawrence Berkeley National Laboratory, Oak Ridge National Laboratory, and Argonne National Laboratory. It featured a unique domestic supply chain including companies and institutes like the National Research Center of Parallel Computer Engineering & Technology, the Shanghai Institute of Applied Physics, and manufacturers in the Yangtze River Delta industrial cluster. Its deployment was part of broader programs akin to Made in China 2025 and national science plans that align with research agendas at Zhejiang University, Fudan University, and the Institute of Computing Technology. The machine was notable in comparisons with systems such as Titan (supercomputer), Sequoia (supercomputer), and Sunway BlueLight.

Architecture and Hardware

The architecture centered on the SW26010 manycore processor created by the Shanghai High Performance IC Design Center and the Sunway Information Industry Company, connecting tens of thousands of nodes via networks inspired by designs from Cray Inc. and ideas circulated in papers from IEEE and ACM. Each node included multiple cores and a hierarchy of on-chip memory, coherent caches, and local scratchpads similar to approaches discussed at SC conferences and in research by teams from University of Illinois Urbana–Champaign, Massachusetts Institute of Technology, and Stanford University. The interconnect topology and system-integration learned from work at Los Alamos National Laboratory and Sandia National Laboratories enabled high-bandwidth communication for parallel workloads used by projects tied to European Centre for Medium-Range Weather Forecasts style climate models and computational fluid dynamics codes from NASA. Storage subsystems were influenced by standards pushed by SNIA and architectures used in installations at National Energy Research Scientific Computing Center.

Performance and Benchmarks

Sunway TaihuLight achieved a LINPACK performance reported at 93.01 petaflops, placing it on the TOP500 list ahead of contemporaries like Titan (supercomputer) and close to efforts at Fujitsu and IBM to push towards exascale. The system's efficiency and sustained performance were analyzed in technical comparisons by teams from Tsinghua University, Peking University, and international benchmarking groups at CERN and Riken. Benchmarks included HPL, HPCG, and domain-specific tests used by researchers at Los Alamos National Laboratory and Argonne National Laboratory for codes such as GROMACS, LAMMPS, and climate model components originating at NOAA. Performance tuning papers appeared at SC and International Supercomputing Conference with authors from Chinese Academy of Sciences and Shanghai Jiao Tong University.

Software and Programming Environment

The programming environment combined a Linux-like operating system developed by Sunway teams and a software stack supporting MPI paradigms similar to implementations from Open MPI and libraries influenced by standards from POSIX and research from University of Cambridge groups. Developers used compilers, runtime systems, and profiling tools analogous to those at Intel Corporation, NVIDIA, and academic projects at University of California, Berkeley to port scientific applications such as ANSYS, OpenFOAM, and community codes like WRF and ACCESS. Optimization strategies drew on work from ACM SIGPLAN and performance engineering groups at Princeton University and Imperial College London. The environment supported parallel I/O, checkpoint/restart frameworks, and middleware comparable to solutions from HDF Group and EPCC.

Applications and Use Cases

Researchers applied Sunway TaihuLight to large-scale simulations in climate science collaborating with agencies like China Meteorological Administration, as well as in computational chemistry and materials science linked to labs at the Chinese Academy of Sciences and industrial partners such as China National Petroleum Corporation for reservoir modeling. It supported engineering workloads for aerospace projects connected to AVIC and automotive research with firms resembling SAIC Motor. Bioinformatics teams from BGI and universities used the system for genomics analyses comparable to workloads run at Broad Institute and Wellcome Sanger Institute. Other uses included seismology studies related to China Earthquake Administration datasets and big data analytics aligned with initiatives at Alibaba Group research labs and Baidu Research.

Development History and Deployment

Development began in earlier iterations of Chinese supercomputing such as the Tianhe series and drew on lessons from systems like Sunway BlueLight and global projects at Riken and Fujitsu that targeted extreme parallelism. The project involved coordination among provincial governments in Jiangsu, national laboratories, and universities including Nanjing University and Southeast University. Deployment and commissioning were announced in 2016 at forums attended by stakeholders from the Ministry of Science and Technology (China), international delegations from organizations like IEEE Computer Society and delegations from research centers including Lawrence Livermore National Laboratory. Subsequent operational management followed models used at National Supercomputing Center of Guangzhou and National Supercomputing Center in Shenzhen with maintenance, software updates, and user allocations overseen by the hosting center in Wuxi.

Category:Supercomputers