China's secretive Sunway Pro CPU quadruples performance over its predecessor, allowing the supercomputer to hit exaflop speeds

Tianhe
(Image credit: Top500.org/News.cn)

Earlier this year, the National Supercomputing Center in Wuxi (an entity blacklisted in the U.S.) launched its new supercomputer based on the enhanced China-designed Sunway SW26010 Pro processors with 384 cores. Sunway's SW26010 Pro CPU not only packs more cores than its non-Pro SW26010 predecessor, but it more than quadrupled FP64 compute throughput due to microarchitectural and system architecture improvements, according to Chips and Cheese. However, while the manycore CPU is good on paper, it has several performance bottlenecks.

The first details of the manycore Sunway SW26010 Pro CPU and supercomputers that use it emerged back in 2021. Now, the company has showcased actual processors and disclosed more details about their architecture and design, which represent a significant leap in performance, recently at SC23. The new CPU is expected to enable China to build high-performance supercomputers based entirely on domestically developed processors. Each Sunway SW26010 Pro has a maximum FP64 throughput of 13.8 TFLOPS, which is massive. For comparison, AMD's 96-core EPYC 9654 has a peak FP64 performance of around 5.4 TFLOPS. 

TOPICS
Anton Shilov
Contributing Writer

Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.

  • bit_user
    Each Sunway SW26010 Pro has a maximum FP64 throughput of 13.8 TFLOPS, which is massive. For comparison, AMD's 96-core EPYC 9654 has a peak FP64 performance of around 5.4 TFLOPS.
    It's a false comparison, though. The SW26010 Pro is a hybrid CPU/GPU. As a hybrid, it packs more raw compute than a CPU, but can't handle general-purpose computation as well. Nor does it have as much raw compute as GPUs, like AMD's MI250X (which packs 28 to 48 fp64 TFLOPS). As such, the best point of comparison is probably with something like Fujitsu's A64FX
    https://www.anandtech.com/show/13258/hot-chips-2018-fujitsu-afx64-arm-core-live-blog
    Another way to think of it is sort of like 6 Cell processors on a chip. Like the Cell, the bulk of its compute lies in the 2-way in-order cores that operate via scratchpad memory. Programming these is probably a lot more like programming a GPU than a CPU. In fact, I wouldn't be surprised if they used OpenCL to utilize them in the exact same way.

    a dual-channel DDR4-3200 (51.2 GB/s) memory subsystem is barely enough for 64 cores, each featuring a 512-bit vector FPU and capable of up to 16 FP64 FLOPS/cycle.
    Big mistake not to use GDDR memory for this.
    Reply