China Plans CPU-Only Exascale Supercomputer with 47,000 Processors

Key Takeaways

• Lingshen would be the first exascale system to hit 2+ ExaFLOPS without GPU accelerators
• The system claims to use only domestic components, but production plans include x86 CPUs with limited Chinese options
• No benchmark data exists since the system hasn't been built, making performance claims unverified

China's National Supercomputing Center in Shenzhen unveiled the Lingshen supercomputer project on April 24. The system aims for sustained performance above 2 ExaFLOPS using only CPUs. No foreign components. No GPU accelerators.

If built as described, Lingshen would pack 47,000 processors into 92 compute cabinets. That would make it the first exascale machine to reach that performance tier without relying on GPUs or accelerator hardware.

A Different Approach to Exascale

Every other exascale system in operation relies heavily on GPUs or accelerator hardware. The U.S. Department of Energy's El Capitan, currently the world's fastest supercomputer, runs on 44,544 AMD MI300A APUs. These chips tightly couple CPU and GPU silicon on a single package.

Lingshen's CPU-only architecture represents a fundamentally different approach. Lu Yutong, director of the Shenzhen supercomputing center and the system's chief designer, presented the technical details at the announcement event.

System Architecture and Specifications

The project splits into two phases: a pilot verification phase and a full production system.

The pilot phase uses 100 Huawei Kunpeng servers built on Arm-based Taishan cores, totaling 12,800 cores. The production system scales dramatically: 1,580 blade servers using x86 CPUs with 101,120 cores and a theoretical peak above 10 petaflops.

• 16 four-way servers adding 2,048 cores
• 4 eight-way servers contributing 1,280 cores
• 36 network cabinets supporting a million-port interconnect
• 650PB of planned storage across 428 nodes
• 67 liquid-cooled storage cabinets with 10 TB/s bandwidth

Performance Claims vs. Reality

Lingshen's claimed sustained performance of 2+ ExaFLOPS would, if achieved, exceed El Capitan's measured Linpack score of 1.809 ExaFLOPS. El Capitan's theoretical peak is 2.79 ExaFLOPS, but real-world Linpack results are always lower than theoretical maximums.

Here's the catch: no Linpack or equivalent benchmark data exists for Lingshen. The system hasn't been built yet. China's claims amount to saying it might achieve 2+ ExaFLOPS at some point in the future.

The Domestic Component Problem

China claims Lingshen will use no foreign-made components. But the production system plans to use x86 CPUs. That creates an obvious contradiction.

China's domestic x86 options are limited. Zhaoxin is a joint venture between VIA Technologies and the Shanghai municipal government. Hygon originally licensed AMD's Zen architecture. Neither has demonstrated the capability to produce processors competitive with current-generation AMD or Intel chips at the scale Lingshen requires.

The pilot phase uses Huawei Kunpeng servers with Arm-based Taishan cores. These are genuinely domestic. But the production system's x86 requirement raises questions about whether "no foreign components" can hold true.

Timeline Uncertainties

The announcement provided no concrete timeline for when Lingshen would become operational. Given the technical challenges and component sourcing questions, it's unclear when or if the system will reach its stated performance targets.

El Capitan already exists and performs at 2.79 ExaFLOPS theoretical peak. For Lingshen to "set a new benchmark for global supercomputing" as China claims, it would need to not only get built but also outperform systems that will likely advance during its development period.

Frequently Asked Questions

What is an exascale supercomputer?

An exascale supercomputer performs at least one exaflop, or one quintillion calculations per second. These systems are used for complex simulations in climate modeling, nuclear research, and AI training.

Why is a CPU-only exascale system significant?

Every current exascale system relies on GPU accelerators for heavy computational work. A CPU-only system reaching 2+ ExaFLOPS would represent a fundamentally different architectural approach.

What is the current fastest supercomputer?

The U.S. Department of Energy's El Capitan holds the top spot with a measured Linpack score of 1.809 ExaFLOPS and a theoretical peak of 2.79 ExaFLOPS.

When will Lingshen be operational?

No concrete timeline was announced. The system hasn't been built yet, and significant questions remain about component sourcing and feasibility.

What processors will Lingshen use?

The pilot phase uses Huawei Kunpeng servers with Arm-based Taishan cores. The production system plans to use x86 CPUs, though China's domestic x86 options are limited.

Source: Latest from Tom's Hardware