The Chinese computer company has removed http://www.hpc100.cn/top100/23/ from its annual list of the top 100 supercomputers. However, this time, the mystery hangs more than ever over the country’s real capacities in terms of high performance computing. The 2024 rankings are virtually a mirror image of last year’s, except for some minor performance improvements that increase total computing power from 1,398 to approximately 1,406 ExaFLOPS.
The first place is taken by a system deployed in 2023. It includes 15,974,400 CPU cores and scores 487.94 PFLOPS in the Linpack test. Although it surpasses the Japanese supercomputer Fugaku (442 FP64 PetaFLOPS), it does not come close to American heavyweights like El Capitan (1,742 ExaFLOPS), Frontier (1,353 ExaFLOPS) and Aurora (1,012 ExaFLOPS).
The second and third systems were launched in 2022, with 208.26 PFLOPS and 125.04 PFLOPS respectively. Chinese authorities’ silence on key details of these cutting-edge systems has led to widespread speculation about their hardware, with some believing they use off-the-shelf CPUs and GPUs obtained through alternative means.
Jack Dongarra, co-founder of Top500.org, previously said that China actually has at least three exascale machines that have not been officially reported. These unlisted systems are said to deliver between 1.3 and 1.7 ExaFLOPS with Chinese-designed hardware, and there’s talk of a 2 ExaFLOPS beast running x86 Hygon processors.
The Tianhe-3 supercomputer – often called “Xingyi” – could be China’s most capable supercomputer. It is rumored to have a peak power of 2.05 exaflops and a sustained power of 1.57 exaflops on Linpack. It is powered by the Matrix-3000 (MT-3000) chip, a blended design that merges general computing and specialized acceleration.
The MT-3000 chip includes 16 CPU cores, 96 control cores and 1,536 accelerator cores, delivering 11.6 teraflops of double-precision power at 1.2 GHz, at a speed of 45.4 gigaflops per watt . This layout is a departure from typically separate CPU-GPU configurations, such as AMD’s MI300A hybrid CPU-GPU approach.
