China’s supercomputer landscape is steeped in intrigue, with the latest Top 100 list offering just a glimpse of its capabilities. Recent updates show slight performance enhancements, but many believe that the rankings only scratch the surface of China’s true prowess in supercomputing. There are whispers of several unreported exascale systems, including the enigmatic Tianhe-3, rumored to employ the advanced Matrix-3000 architecture.
The Chinese Society of Computer Science has unveiled its annual ranking of the nation’s Top 100 supercomputers, and the 2024 list is almost a carbon copy of last year’s roster, save for minor upgrades bumping the total computing power from 1.398 to approximately 1.406 ExaFLOPS. Topping the list is a system launched in 2023, boasting 15,974,400 CPU cores and achieving 487.94 PFLOPS in the Linpack test. Despite outpacing Japan’s Fugaku, it doesn’t quite match up to the powerhouse American systems like El Capitan, Frontier, and Aurora.
The second and third systems on the list emerged in 2022, with performances of 208.26 PFLOPS and 125.04 PFLOPS, respectively. However, the limited information provided by Chinese authorities about these leading machines leaves room for speculation regarding their hardware, leading some to suspect they use CPUs and GPUs acquired by unconventional means.
Adding to the mystery, Jack Dongarra, a co-founder of Top500.org, has posited that China could have at least three unreported exascale systems in operation. Reports suggest these covert machines deliver between 1.3 and 1.7 ExaFLOPS and run on technology developed within China. There is even talk of an elusive 2 ExaFLOPS supercomputer powered by Hygon x86 processors.
Among these rumored giants, the Tianhe-3, known as “Xingyi,” stands out as potentially China’s most powerful supercomputer. Boasting a 2.05 exaflops peak and 1.57 exaflops of sustained performance on Linpack, it is powered by the innovative Matrix-3000 (MT-3000) chip. This chip employs a groundbreaking mixed design, combining general-purpose computing with specialized acceleration, featuring 16 CPU cores, 96 control cores, and 1,536 accelerator cores. Offering 11.6 teraflops of double-precision power at 1.2 GHz and an efficiency of 45.4 gigaflops per watt, it represents a departure from conventional separate CPU-GPU configurations, akin to AMD’s MI300A hybrid approach.
This blend of cutting-edge design and strategic secrecy continues to make China’s supercomputing sector a subject of global fascination, hinting at even greater advances yet to be disclosed.
