What Drives China’s Emerging CPU Landscape? A Deep Dive into Six Domestic Chipmakers

Background

The CPU is a core component of hardware infrastructure. The dominant global architectures are ARM and x86, both largely foreign‑designed, resulting in a low domestic substitution rate. China launched the Taishan Program and the 863 Program to develop indigenous CPUs, and in 2006 the National High‑Performance Computing and Chip Initiative ("Nuclear‑High‑Base") further accelerated support for domestic designs.

1. Overview of Domestic CPU Vendors

HaiGuang and Zhaoxin use licensed x86 IP cores, allowing them to start from a high‑performance baseline and benefit from a mature ecosystem, but they must pay licensing fees and have limited innovation freedom. HaiGuang’s latest CPUs approach the performance of high‑end international products and remain fully compatible with the x86 instruction set.

Kunpeng (Huawei) and Phytium (Tianjin) license the ARM instruction set, enabling them to design custom cores and System‑on‑Chip (SoC) solutions with higher autonomy. Huawei’s Kunpeng 920 is the industry’s first 7 nm ARM server processor, delivering leading compute density and performance comparable to top x86 chips.

Loongson (Institute of Computing Technology) developed the proprietary LoongArch ISA, offering strong independence and an open‑collaboration model that supports IP licensing, chip‑level development, and system‑kernel integration.

ShenWei created its own 64‑bit instruction set for special‑purpose domains, focusing on defense and network‑security applications. Its ecosystem has matured alongside its products.

RISC‑V has gained rapid traction due to its concise ISA and permissive BSD‑style license; Chinese firms such as Alibaba’s Pingtouge and Guoxin Technology have released RISC‑V‑based products. MIPS and Alpha remain niche, represented domestically by Loongson (MIPS‑compatible) and ShenWei (Alpha‑derived).

2. Kunpeng 920 Processor

The Kunpeng 920 supports the ARMv8.2 ISA and is the first 7 nm ARM data‑center processor, optimized for big‑data and distributed‑storage workloads. It features out‑of‑order execution, multiple issue, and advanced branch prediction to boost single‑core performance.

Key specifications:

64 cores

8‑channel DDR4 memory interface

Base frequency up to 2.6 GHz

Peak memory bandwidth 1.5 TB/s

PCIe 4.0 and CCIX support, total I/O bandwidth 640 Gbps

Huawei Cache Coherence System (HCCS) 480 Gbps inter‑chip link, supporting up to four Kunpeng 920 CPUs and a 256‑core NUMA configuration

Performance comparisons show that at 48 cores the Kunpeng 920 matches Intel Xeon 8180 in integer workloads while consuming 20 % less power; at 64 cores it exceeds Xeon 8180 by roughly 33 %.

3. Phytium Processors

Phytium offers three product lines:

High‑performance server CPUs ("Tengyun S" series)

High‑performance desktop CPUs ("Tengyun D" series)

High‑end embedded CPUs ("Tengyun E" series)

All lines are based on ARM licenses (including a permanent ARMv8 license) and provide customized solutions for various industry verticals.

4. HaiGuang Processors

HaiGuang partners closely with AMD, licensing the x86 core and Zen architecture. AMD receives a US$293 million royalty. HaiGuang builds its CPUs on this IP, allowing it to bypass Intel’s licensing restrictions while delivering a domestically produced x86 product that aligns with national policy.

Benchmarking against six Intel Xeon Platinum 2020 models shows HaiGuang’s latest 7285 CPU approaches the performance of high‑end international counterparts in typical workloads.

5. Zhaoxin "KX" and "KH" Processors

Zhaoxin’s KX‑6000 and KH‑30000 series integrate CPU, GPU, and chipset on a single SoC, simplifying system development and reducing cost. The KX‑U6780A (8‑core) still lags behind Intel i5‑7th‑gen in overall performance, especially single‑core speed, but it surpasses the i5 in integer benchmarks.

6. Loongson LoongArch ISA and 3A/3B Series

LoongArch offers extensive extensibility with roughly 2 000 instructions covering base architecture, vector operations, virtualization, and binary translation. The 3A5000/3B5000 series target PCs and servers, using the LA464 micro‑architecture and higher clock rates to improve performance and reduce power consumption.

SPEC CPU2006 results for the 3A5000:

Single‑core integer score: 25.1

Single‑core floating‑point score: 26

These scores are comparable to domestic ARM V8 4‑core 7 nm chips and exceed ARM V8 4‑core scores by ~10 % (integer) and ~100 % (floating‑point). Multi‑threaded performance remains behind Intel i5‑9500 but ahead of domestic ARM V8 4‑core solutions.

7. ShenWei Alpha‑Based Processors

ShenWei originally extended the DEC Alpha ISA to create a proprietary 64‑bit instruction set. The first ShenWei CPU (SW‑1) was fabricated in 2006 on a 130 nm process at 900 MHz. Subsequent generations (SW‑26010, SW‑831, SW‑4E, etc.) introduced out‑of‑order execution, SMT, and advanced pipeline designs, achieving double‑digit performance gains while maintaining power efficiency.

The SW‑26010 processor delivers over 12.5 × 10^15 floating‑point operations per second, ranking among the world’s fastest super‑computing chips. It powered the Sunway TaihuLight system, which topped the TOP500 list with a peak performance of 12.54 × 10^18 FLOPS and a sustained performance of 9.3 × 10^18 FLOPS.

Conclusion

China’s domestic CPU ecosystem spans a wide range of architectures and market segments, from x86‑licensed high‑performance servers to fully indigenous LoongArch and RISC‑V designs. National programs and strategic partnerships have accelerated development, resulting in several products that now rival international flagship CPUs in specific benchmarks, while also highlighting remaining gaps in single‑core performance and ecosystem maturity.