- Lingsheng system targets two exaFLOPS using only central processing units
- Unique CPU architecture challenges GPU-dominated supercomputing industry standards
- System design integrates high-bandwidth memory and high-speed interconnection networks
A Chinese supercomputing center has announced plans for a machine that would reach two exaFLOPS using nothing more than central processors.
The Lingsheng system, unveiled at an April 2026 conference in Shenzhen, would include 47,000 processors in just 92 computing cabinets.
Lu Yutong, director of the National Supercomputing Center in Shenzhen and chief designer of the system, explained that the hardware and software stack is “completely controllable independently.”
Article continues below.
A fundamentally different architectural strategy
Exascale machines in the industry currently rely heavily on GPU accelerators or specialized hardware.
This makes the CPU-only approach a significant departure from established global trends.
The system leverages domestically produced high-performance CPUs along with high-bandwidth on-chip memory and high-speed interconnect networks.
It also incorporates 3D floating orthogonal computing and full liquid cooling to manage thermal output.
According to the announcement, the Lingsheng platform makes progress in six major technical areas: architecture, performance, power consumption, scheduleability, scalability and reliability.
The system supports exascale computing power with exascale storage and petascale communication, and employs what officials described as the world’s largest-scale centralized liquid cooling technology.
A pilot verification phase uses 100 Huawei Kunpeng servers built on Arm-based Taishan cores, with a total of 12,800 cores.
When scaled up to full production, the same system design would incorporate 1,580 blade servers using x86 CPUs with 101,120 cores and a theoretical peak greater than 10 petaflops.
The entire infrastructure also features 36 network cabinets that support one million port interconnection.
It will also feature 650 PB of planned storage across 428 nodes and 67 liquid-cooled storage cabinets offering 10 TB/s of bandwidth.
The world’s fastest computer today, the US Department of Energy’s El Capitan, is powered by 44,544 AMD MI300A APUs, integrating CPU and GPU silicon into a single package.
If Lingsheng’s sustained performance of 2 exaFLOPS is achieved, it will surpass El Capitan’s measured Linpack score of 1,809 exaFLOPS.
On the other hand, the figure of 2 exaFLOPS for the Lingsheng system is a theoretical number, but El Capitan already has a theoretical value of 2.79 exaFLOPS.
Therefore, the claim of surpassing the fastest computer in the world does not seem achievable if theoretical values are compared with each other.
Unanswered questions and untested capabilities
Several critical questions remain unanswered regarding the Lingsheng system, primarily because no reference data exists for the machine.
Although China claims that this system will not rely on non-Chinese suppliers, the country’s domestic x86 options remain limited to Zhaoxin and Hygon.
None of these national alternatives have proven to have processors that can compete with current generation parts from Intel or AMD.
The announcement also did not name specific suppliers for the production system and did not provide any operational timeline for its completion.
In terms of potential applications, the technology covers nine fields, including remote sensing, materials science, bioinformatics, meteorology, pharmaceutical industry, oil exploration, artificial intelligence, life sciences and electromagnetic simulation.
One research team reported achieving 81% parallel scalability for first-principles calculations involving 100 million atoms.
Another group claimed that virtual screening of compounds on a trillion-scale could improve efficiency 1,000-fold through a combination of artificial intelligence and reinforcement learning.
However, these remain theoretical statements until a working machine produces verifiable reference results.
Through Tom Hardware
Follow TechRadar on Google News and add us as a preferred source to receive news, reviews and opinions from our experts in your feeds.
