Tech news in 3 minutes

This is the Storage of Spaceborne Computer 4 Bringing AI Compute to the Moon

27 d ago

The Chinese LineShine supercomputer, developed by the National Supercomputing Centre in Shenzhen (NSCS), has claimed the #1 spot on the latest Top500 list, becoming the first system to publicly exceed 2 ExaFLOPS on Linpack with a measured performance of 2.2 EF—22% faster than the previous leader, El Capitan. Notably, LineShine is a pure CPU system, unlike its GPU-based competitors. It is built around the LX2 CPU, believed to be designed by Huawei, which integrates two compute dies with 152 Armv9 cores each (304 cores per CPU), featuring SVE and SME vector/matrix accelerators. Each CPU includes 32GB of HBM and 256GB of DDR5, consuming 690 watts, and has integrated 800Gbps NICs. Two CPUs form a node, eight nodes per blade, 16 blades per compute frame (256 CPUs), two frames per cabinet (512 CPUs), with approximately 89–90 cabinets totaling around 13.79 million cores. The system uses a LingQi 1.6Tbps network in a 4-layer fat-tree topology. LineShine achieves 80% of its theoretical peak (2.74 EF) on Linpack, outperforming US exascale systems (50–65% efficiency). However, it consumes 42MW of power (52 GFLOPs/watt), slightly less efficient than El Capitan (60 GFLOPs/watt), and is liquid-cooled. The Top500 list saw major shifts: AMD-powered systems (El Capitan and Frontier) drop to #2 and #3, while AMD now powers 4 of the top 10. NVIDIA holds three Hopper-based spots (#5, #7, #10), Intel’s Aurora is #4, and the Italian Eni HPC7 (AMD MI300A) debuts at #6. Japan’s Fugaku (Armv8) falls to #9. The article notes that many large systems are not benchmarked for Top500, and AI hardware is reducing FP64 support.