Top500: China Opts Out of Global Supercomputer Race (2024)

The U.S. has two of the world’s fastest computers, highlighting significant hardware progress in the country, according to the latest Top500 list of fastest supercomputers released on Monday.

But China — once home to the world’s fastest supercomputers — has opted to remain quiet about its advances in supercomputing.

China already has powerful exascale systems, but it is shielding the systems from rivals looking to get a peek into its secretive technology infrastructure.

The Top500 list recognizes 500 of the world’s fastest computers based on benchmarks stipulated by the organization. The Top500 list is highly influential, but China — once an active participant — has lost interest in it.

The top two systems in the latest list are installed at U.S. government labs and are the only publicly known computers to pass 1-exaflops in total performance.

China’s exascale supercomputers may be as powerful as the fastest U.S. systems. The country hasn’t shared information about the systems with the Top500 organizers.

“In fact, China did not report a single new machine for this new list,” the Top500 organizers said in a release.

The U.S., with 168 systems (up from 161 from the November list), and China, with 80 (down from 104), had the most entries on the latest Top500 list.

New Age Cold War

The U.S. has restricted exports of high-performance chips, including the latest AI chips — including cutting-edge NVIDIA GPUs — to China. The goal is to cripple China’s advances in AI, which require the latest chips.

The U.S. has also banned the export of tools from companies ASML that are used to make advanced mobile, AI, PC and server chips. But China’s Huawei recently released a smartphone with a 7-nanometer chip, which was considered unachievable by the U.S. in the wake of the restrictions.

Chinese semiconductor manufacturing firm SMIC is now reportedly making chips using the 5-nm process, which is a generation behind the latest chips made by Taiwan Semiconductor Manufacturing Co.

SMIC’s chips are made using old manufacturing tools, and there are doubts on China advancing beyond that node.

The trade restrictions imposed by the U.S. are forcing China to focus on tech self-sufficiency. The country is pouring billions into hardware development.

In March, China banned the use of x86 chips from AMD and Intel in government PCs. China is mandating PCs use homegrown chips. The existing homegrown chips, including Loongson, can’t achieve the cutting-edge performance of the latest Intel or AMD processors.

China is also pouring money into the RISC-V instruction set architecture, which is an alternative to x86 and ARM architectures.

RISC-V chips are still far off from maturity, but hundreds of Chinese companies are developing RISC-V CPUs for mobile devices, PCs and servers. Notably, Alibaba has released open source server RISC-V CPU designs.

The Fastest Computers Today

Supercomputers are critical to the national and economic interests of countries. These systems are used to develop weapons, maintain nuclear weapon stockpiles, and make economic projections.

The fastest supercomputer today, called Frontier, is hosted at the U.S Department of Energy’s Oak Ridge National Laboratory. It runs on AMD’s 3rd Gen Epyc CPUs and MI250X GPUs, and peaks at 1.2 exaflops of performance.

The second fastest supercomputer, Aurora, is hosted at DOE’s Argonne National Laboratory. The system runs on Intel’s CPUs and GPUs, and is only partially benchmarked.

Aurora can peak at 2-exaflops, and the system could topple Frontier from the top spot when full benchmarks are complete. But the system consumes a significant amount of power.

The supercomputing benchmarks, typically known as high-performance Linpack, require plenty of fine-tuning until the performance thresholds are achieved.

The U.S. government stipulates that the supercomputers meet the specified benchmark requirements when handing out contracts to system builders. HPE is the contractor for Frontier and Aurora.

China has provided a peek into its exascale systems via research papers. The Next Platform has also unearthed some details about the systems.

A Chinese research project on a 1.5-exaflops system was considered for the highest supercomputing honor, the Gordon Bell prize, last year.

Minus the exascale systems, the fastest Chinese system on the Top500 list is Taihulight, which is currently ranked 13th. The system maxes out at 93 petaflops, which is more than 10 times slower than the exaflop-scale supercomputers today.

TaihuLight was the world’s fastest system in 2016 and 2017, until it was toppled by ORNL’s Summit system — which has IBM’s Power9 chip — in 2018.

The second-fastest Chinese system, Tianhe-2A, is ranked number 16 on the Top500 list. It was the world’s fastest supercomputer from 2013-2016.

Concern for TOP500

The lack of submissions from China has put Top500 organizers in a bind. The list no longer provides full visibility on the fastest systems in the world.

However, the organizers have said they are working with their Chinese counterparts to get details on the types of systems in use.

The Top500 list is aging fast, and the number of new systems has declined since 2017. The average age of Top500 systems is now 30 months, compared to an average age of 15 months in 2018-2019.

Newer systems are focused on AI, with faster memory and throughput. System builders are also more focused on the power efficiency of systems than performance.

Cloud makers, including Google, Microsoft and Amazon, are pouring billions into building their computing infrastructure around AI, which uses mixed precision techniques for deterministic computing, which is the default for AI-style computing.

Upcoming Supercomputers

Anticipation is growing for the 2-exaflop supercomputer called “El Capitan,” which is already under construction at the Lawrence Livermore National Laboratory. It could make its first appearance in the next Top500 list in November.

El Capitan has AMD’s CPUs and upgraded MI300 GPUs, which compete with Nvidia’s GPUs. The supercomputer is designed for AI and high-precision scientific applications.

Upcoming exascale supercomputers in Europe, called Jupiter and Jules Vernes, have modular designs where quantum and AI computing modules can be quickly plugged into the infrastructure as and when needed.