The supercomputer Tianhe-3 is believed to be the fastest in the world due to its utilization of the new processor MT-3000, designed by the National University of Defense Technology (NUDT), specialized in developing supercomputers and the Chinese space program.
A recent article from The Next Platform unveiled details about the MT-3000 processor, featuring a unique design that combines CPU cores, control cores, and accelerator cores.
According to the website, the MT-3000 processor has a multi-design architecture with 16 cores for the CPU unit, 96 control cores, and 1536 accelerator cores.
The MT-3000 processor is capable of achieving a performance of up to 11.6 trillion calculations per second, with an energy efficiency of up to 45.4 gigaflops/watt, and a frequency of up to 1.20 gigahertz.
One advantage of the MT-3000 processor design is that it integrates CPU cores, control cores, and accelerator cores in the same silicon chip, partially reflecting the design philosophy of the hybrid AMD Instinct MI300A processor; indicating a more cohesive and efficient design approach rather than using separate CPU and GPU processing systems. Compared to the multi-chip AMD MI300A processor, the MT-3000 processor appears to rely on a single-chip design.
It is noteworthy that the MT-3000 processor requires manufacturing precision of up to 10 nanometers or 7 nanometers, due to the high core density in the processor architecture.
The manufacturing company of the processor is believed to be the Chinese SMIC, which also produces the latest HiSilicon Kirin 9000S processor for Huawei using a 7-nanometer manufacturing accuracy in the second generation.
However, it is not clear if SMIC has the sufficient production capacity with 7-nanometer manufacturing accuracy to meet the needs of both Huawei and NUDT. As a result, the MT-3000 processor may be used with a different manufacturing precision.
The Tianhe-3 supercomputer is expected to achieve unprecedented computational performance in the world thanks to the MT-3000 processor it relies on, with an expected capability of reaching 1.57 exaflops of computational operations per second during LINPACK tests.
In comparison, the Frontier supercomputer, considered the fastest supercomputer in the United States, achieves performance of up to 1.102 exaflops of computational operations per second.
