OceanLight, the Chinese exascale supercomputer that has defied listing in the Top500, has once again surged to prominence as a finalist in the Gordon Bell Prize. A yearly award in recognition for "outstanding achievement in high-performance computing," the Gordon bell Prize awards a $10,000 fund to the winner, picked from a group of five finalists. OceanLight, operated by the University of Science and Technology of China, once again races neck and neck with crowning jewels of Western supercomputing achievement such as Frontier and Summit, as well as Japan's Arm-based Fugaku.
OceanLight's supercomputing chops are impressive, especially when you take into account its technology. Besides being an exascale-class supercomputer, one of only a few in the world right now, OceanLight doesn't have the same amenities as those created outside of heavily sanctioned China. For one, China is barred from accessing cutting-edge manufacturing tech. This makes it wholly dependent on its internal manufacturing capabilities to provide most of the required components and technology that go into building such a massive machine.
This means that OceanLight has been built with technologies that have mostly been deprecated compared to that of the current Best GPUs and Best CPUs enthusiasts can get their hands on, let alone the cutting-edge manufacturing processes and technologies powering the likes of El Capitán, which will employ AMD-made MI300 APUs. Semiconductor Manufacturing International Corporation (SMIC), China's homegrown CPU manufacturer, built OceanLight's exascale capabilities by leveraging a 14nm process for its SW26010-Pro CPUs.
According to some very educated math from Nextplatform, each SW26010-Pro CPU can deliver 14.03 teraflops of compute at either FP64 or FP32 precision and 55.3 teraflops at BF16 or FP16 precision. Extracting data from the previous Gordon Bell Prize entry the system took part in, it's estimated that OceanLight features a 107,520 node architecture with one SW26010-Pro per node, for a total of 41.93 million cores.
As each cabinet can support 40 nodes, that works out to 105 cabinets in total. And according to this year's Gordon Bell entry, the supercomputing workload that simulated as many as 2.5 billion atoms and their interactions called as many as 28.1 million of OceanLight's cores.
China's leveraging whatever technology it can to advance both its quantum and supercomputing capabilities, and the country isn't shy of the increased power consumption (mostly coal-based) and materials requirement that comes from scaling older, less advanced technology. Once again, where there's will and state-backed pockets, there's a way to plow forward
US requests proposals for next-gen Discovery supercomputer — will be up to five times faster than the world's fastest supercomputer, arrive in 2027
Thunderbird packs up to 6,144 CPU cores into a single AI accelerator and scales up to 360,000 cores — InspireSemi's RISC-V 'supercomputer-cluster-on-a-chip' touts higher performance than Nvidia GPUs
