Direction China, which has announced a new supercomputer project with LineShine. This new beast is announced to offer a total computing power of over 2 ExaFlop. Better (or worse, depending on your point of view), components will be sourced 100% locally, ensuring the project’s total independence from outside players. You’ll see, the datasheet is insane!
LineShine: a supercomputer with over 2 ExaFlop of capacity!
An insane datasheet!
We’re talking about a supercomputer based mainly on CPUs , since the configuration presented would bring together 47,000 processors. All of this will be distributed across 92 computing cabinets. The industrial complex will bring together a total of 1,580 blade servers, all equipped with domestic processors, i.e. a total of 101,120 cores. Cooling is claimed to be the world’s largest liquid cooling system (integrated in 67 cabinets), with 3,214m of pipe and a weight of 243.9t. There’s plenty of room for storage, too, with 650 PB of capacity. As for bandwidth, we’re talking about 10Tb/s.
Let’s not forget to mention the 36 cabinets dedicated to the network that will be used to interconnect this little world.
In the meantime, a pilot system dedicated to verifying the architecture will employ a total of 100 Huawei Kuanpeng servers integrating a total of 12,800 cores.
A 100% local project!
But the most interesting aspect of this project lies in the way components will be sourced. Indeed, LineShine will be a 100% Chinese supercomputer, with components produced in China. As a result, the country doesn’t have to worry about anyone’s mood swings, or any sanctions that might hit the country at an inopportune moment.
The project is ambitious, since it aims to dethrone El Capitan, the most powerful supercomputer currently available, based on AMD hardware with a maximum power of 2.8 ExaFlop. The Chinese model plans to support the FP64/FP32/FP16 and Int8 software stack. It is also designed to be easy to use, and will integrate a complete software ecosystem featuring a compiler, debugger and more. In terms of AI performance, a single CPU generates 578 token/s on DeepSeek. Once complete, we expect throughput to be 100x higher.
Finally, LineShine will be used to run AI models, but not only. It is also intended for weather simulation, bio informatics, materials simulation, etc.
