KEY POINTS
- LineShine, a Chinese supercomputer built only on CPUs, leads the new Top500 with 2.2 exaFLOPS in the High Performance Linpack benchmark.
- The US Department of Energy is planning at least nine new systems under its Genesis Mission, most of them smaller or AI-oriented.
- The Register argues that Linpack is losing importance as a prestige benchmark as AI automation of research becomes the new focus.
China holds the top position on the newest Top500 list of supercomputers, with little changing at the head of the ranking, The Register reported. The National Supercomputing Centre in Shenzhen runs the LineShine system, which reached 2.2 exaFLOPS in the High Performance Linpack (HPL) benchmark.
That result is about 22 percent higher than El Capitan at Lawrence Livermore National Laboratory, the largest US machine on the list. LineShine draws 42.2 megawatts, so it may not be the most efficient system, but The Register notes that the top rank is not awarded for efficiency. It also points out that the machine was built despite US export controls that block Chinese buyers from American GPUs.
LineShine has no GPUs at all. It uses 45,360 LingKun LX2 processors, each with 304 Arm v9.2 cores running at 1.5 GHz, spread across 22,680 dual-socket nodes. According to the report, each LX2 carries 16 HBM memory stacks and enough vector and matrix hardware that it resembles a GPU, yet it is in practice a very large CPU. The Register says this is the first time in almost a decade that China has led the list.
The US is unlikely to regain first place soon, even though the Department of Energy is working on it. Under the Genesis Mission, announced late last year, the department expects to deploy at least nine new supercomputers, but most will be smaller or built mainly for AI. One example is Lux at Oak Ridge National Laboratory, built with HPE and AMD, which delivers 220 petaFLOPS in HPL and ranks 14th. At FP4 precision its peak exceeds 20 exaFLOPS, a figure suited to AI inference and some training.
Oak Ridge's upcoming Discovery system could challenge LineShine. Details are limited, but it is an HPE Cray machine that will combine AMD Venice CPUs with Instinct MI430X GPUs, which retain strong double-precision performance. The Department of Energy says Discovery will be three to five times faster than Frontier, which The Register converts into a theoretical FP64 peak of 5.1 to 8.5 exaFLOPS.
The outlet also notes that the US has private AI clusters that could beat the leaders if their operators chose to run HPL. It cites SpaceXAI's 300-megawatt Colossus in Memphis, Tennessee, with 200,000 Hopper GPUs. Even at an assumed 25 percent efficiency, the outlet estimates it could exceed three exaFLOPS. Such operators have little reason to do so, it says, because renting the hardware for AI training and inference is more lucrative than a benchmark ranking.
The Register's broader argument is that the contest has moved from raw simulation power to AI models and agents that can automate laboratories and test hypotheses around the clock. That, it says, explains the Genesis Mission's AI emphasis. Supercomputing and the Top500 stay relevant as long as more FLOPS means more simulations, but the list no longer shows the largest computers ever built.