How a Zero‑GPU, All‑CPU Supercomputer Defied the GPU Trend to Claim World #1

At the ISC2026 conference in Hamburg, the TOP500 list announced that the Chinese supercomputer “Ling Sheng” reclaimed the world‑first position with a sustained double‑precision performance of 2.19 EFLOPS, the first system to surpass the 2 EFLOPS barrier and marking China’s return to the summit after nine years.

The latest TOP500 shows a clear trend: GPU acceleration now dominates high‑performance computing, with 55.4% of systems equipped with GPU accelerators from AMD or NVIDIA. In contrast, Ling Sheng follows a completely different, non‑mainstream route.

According to system chief Lu Yutong, Ling Sheng pioneered an “Online Acceleration” all‑CPU architecture that embeds an AI matrix accelerator directly inside the CPU, eliminating the need for separate GPU cards. This choice is partly driven by ongoing export bans on AMD and NVIDIA GPUs to China, showcasing a fully domestic, stack‑level capability.

The design philosophy argues that instead of shuttling data between CPU and GPU, deep integration of compute and AI at the chip level yields better performance. Jack Dongarra, Turing Award laureate, highlighted that Ling Sheng offers a hopeful glimpse of a new AI‑for‑Science system architecture.

Beyond the CPU innovation, the system achieves full‑stack breakthroughs: the self‑developed LX2 CPU integrates the first domestic HBM, delivering a ten‑fold memory‑bandwidth increase; the LingQi high‑speed interconnect supports two million ports and 100 k nodes for massive scale‑out; and a custom software stack exposes hardware capabilities in a programmable, optimizable manner.

Energy efficiency is also a highlight: Ling Sheng ranks 50th on the GREEN500 list with a power efficiency of 51 GFlops/W and employs 100% liquid cooling.

Performance-wise, the system maintains an average scaling efficiency of 84.4% when running on more than ten million cores, enabling efficient execution of large‑scale scientific and AI workloads such as climate modeling, engineering simulation, materials science, drug discovery, brain research, scientific AI, and large‑model inference.

The article notes that China’s previous TOP500 victory was the “Sunway TaihuLight” in 2017; after years of export restrictions, Ling Sheng breaks the long‑standing US/European dominance, illustrating that multiple technical paths can lead to leadership and that China can now be a definition‑setter in the evolving era of supercomputing.