Zhaoxin KH-4000: Chinese 16-core chip rivals AMD EPYC

Zhaoxin KH-4000: Chinese 16-core chip rivals AMD EPYC

A 16-core processor running at 2.7 GHz, manufactured on a 16-nanometer process and without SMT, should not be able to compete with an AMD EPYC 7601. Yet, the Zhaoxin KH-4000 does exactly that. An analysis published on June 22, 2022, places this Chinese chip on par with the first-generation EPYC under SPEC CPU 2006: it loses slightly in some cases and wins narrowly in others.

The Chinese CPU antiestéticatured in this news story is not aiming to win the home PC battle. Its target is servers, and that is where the crucial detail lies: Beijing's goal is not to challenge Intel and AMD for market share, but to stop buying from them. When that happens, the issue ceases to be commercial.

What is the Zhaoxin KH-4000 and what’s inside?

16 cores without SMT — meaning 16 threads —, 2.7 GHz frequency, DDR4 memory support, PCIe 3.0, and 32 MB of L3 cache. In a dual-socket configuration, the chip totals 32 cores. The manufacturing node is 16 nm, and its TDP is rumored to be around 100 W.

On paper, nothing spectacular. The nuance appears when looking sideways at the rival: despite using a lithographic process inferior to Intel’s, the Chinese manufacturer integrates more L3 cache than the Xeon Platinum 8153, which has only 22 MB. Integrating more fast memory into the chip is not just a brochure figure: it prevents cores from spending all day waiting for main memory.

The compared EPYC 7601 doesn’t play in the same league

The comparison is ungenerous, and those who made it admit so. The AMD EPYC 7601 antiestéticatures 32 cores and 64 threads, boosts up to 3.2 GHz, and declares a TDP of 180 W. It doubles the Zhaoxin in core count, yet the Chinese chip holds its own in the benchmark suite.

Translated: in raw power, the EPYC dominates; in efficiency per watt invested, the advantage narrows until it disappears. That a newcomer with 16 nm can match a first-batch 14 nm processor says more about the Chinese manufacturer than about the American processor.

100 watts vs. 125 and 205: efficiency as the argument

Intel’s numbers set the scale. The Xeon Platinum 8153, with 16 cores and 32 threads at 2.8 GHz, declares 125 W. The Xeon Platinum 8180M, with 28 cores and 56 threads at 3.8 GHz, goes up to 205 W. The Zhaoxin, with worse lithography, stays below both.

This is where the analysis stops being a curiosity: in a data center, consumption is paid monthly, and in a country deploying thousands of servers for its own administration, that differential is worth more than any medal in a benchmark.

Why doubt the performance of the Chinese CPU?

Because the design isn’t built from scratch. Skepticism points to these processors deriving from VIA patents and its subsidiary Centaur Technology —with Cyrix lineage in the background— and that such a leap doesn’t fit without more data. The request is prudent: cross-check with Chinese sources before accepting the result as valid.

The industry narrative itself carries this baggage. Those who lived through the nineties remember the Cyrix 6x86 at 166 MHz from 1996 as a stable and cheap machine that never threatened anyone. The question is whether the Zhaoxin is a reissue of that or the first serious warning.

IPC, compilers, and the bottleneck that doesn’t always exist

Performance per cycle is the king metric in comparisons, and also the most debated. A poorly parallelized program performs the same with 8 or 100 cores; a compiler optimized for a specific architecture can shift the result of an entire suite by several points without touching the silicon.

In closed environments, without the burden of backward compatibility, that discussion becomes blurrier. The Itanium case —an architecture with its own compiler and promise of domestic range— is cited as a reminder that the best design doesn’t win if the market turns its back. The repeated conclusion: more tests are needed, and doubt is reasonable in both directions.

From benchmark to boardroom: sanctions and deglobalization

The core issue isn’t technical. For decades, manufacturing in China was the cheap route, and with each offshored plant traveled knowledge of how things are done. The result, argues part of the analysis, is that the country no longer needs the West to take the next step.

The other side is security. It is argued that Western CPUs include instructions not found in public documentation, and replacing them with proprietary silicon is also a matter of digital sovereignty. The calculation is cold: whoever controls the processor controls what runs on top of it.

That is the point nobody closes. Some see in the Zhaoxin only a rebranding of foreign designs twenty years behind; others see the first brick of an industry that won’t stop.

That a 16 nm chip with half the cores ties with an EPYC and consumes half doesn’t close any debate, it opens one: the announced next step —7 nm, DDR5, and PCIe 4.0— has no date, and precisely because of that, it deserves to be on the radar.

---

Notice: comparative performance figures come from the analysis published in June 2022 and have not been independently replicated.