Chinese AI chip, Huawei Ascend, Cambricon, Hygon, NVIDIA alternative China
The global AI chip market has long been dominated by a single name: NVIDIA. But as U.S. export controls tighten and geopolitical tensions reshape the semiconductor landscape, a new generation of Chinese AI chips is emerging to challenge that dominance. Huawei's Ascend series, Cambricon's思元 (Siyuan) lineup, and Hygon's DCU accelerators are at the forefront of China's push for AI silicon self-sufficiency.
This article examines where each of these Chinese AI chipmakers stands today, how their hardware compares to NVIDIA's A100 and H100, what ecosystem gaps remain, and what the 2026–2030追赶 timeline looks like.
The Strategic Context: Why China Needs Its Own AI Chips
Since 2022, the U.S. Bureau of Industry and Security (BIS) has progressively restricted China's access to advanced GPUs. The A100 and H100 were barred from export to China in September 2022 [1]. The pared-down A800 and H800 workarounds were subsequently blocked in October 2023 [2]. By 2024, even the RTX 4090 consumer GPU faced export restrictions.
For Chinese cloud providers, research institutions, and AI startups, the message was clear: building a sovereign AI compute stack is no longer optional—it's an existential imperative.
The result has been an unprecedented wave of domestic chip development. Three companies have emerged as the most credible challengers to NVIDIA's throne in China.
Huawei Ascend: The Flagship Challenger
Ascend 910B: Closing the Gap
Huawei's Ascend 910B, launched in late 2023, represents the most serious Chinese attempt to match NVIDIA's data center GPUs. Built on a 7nm process—reportedly fabricated by SMIC using DUV lithography despite U.S. sanctions on EUV equipment—the 910B delivers approximately 376 TFLOPS of FP16 performance and 640 TFOPS of INT8 [3].
That puts it in the same ballpark as the NVIDIA A100's FP16 Tensor Core performance of 312 TFLOPS, though the A100 benefits from superior memory bandwidth (1,555 GB/s HBM2e vs. the 910B's estimated 1,200 GB/s).
Ascend 910C: The Next Leap
The Ascend 910C, expected to enter volume production in 2026, is designed to target H100-class performance. Industry reports suggest it aims for over 800 TFLOPS of FP16 compute, potentially leveraging advanced packaging and architectural refinements [4]. However, the persistent lack of EUV lithography equipment remains a fundamental constraint on process node advancement.
Huawei's key advantage is vertical integration. Through its HarmonyOS ecosystem, Kunpeng server CPUs, and MindSpore AI framework, Huawei controls the full stack—from silicon to application layer. This is China's closest analog to NVIDIA's CUDA-to-GPU vertical integration.
Cambricon: The Pure-Play AI Chip Specialist
Cambricon Technologies (寒武纪), spun out of the Chinese Academy of Sciences in 2016, is China's only listed pure-play AI chip company. Its思元 (Siyuan) series has evolved through several generations:
- Siyuan 290 (2019): China's first cloud AI training chip, 7nm, up to 128 TOPS INT8 - Siyuan 370 (2021): Improved training performance, adopted by China Telecom and several national AI projects - Siyuan 590 (2024–2025): Cambricon's most advanced chip to date, targeting general-purpose AI training and inference workloads [5]
Cambricon's strength lies in its focused R&D. Unlike Huawei, which operates across telecom, consumer electronics, and enterprise infrastructure, Cambricon does one thing: AI chips. This focus has produced deep expertise in accelerator microarchitecture, particularly in sparsity exploitation and mixed-precision compute.
However, Cambricon's weakness is scale. The company has struggled with commercial adoption outside of state-backed projects. Without a dominant cloud platform of its own (unlike Huawei Cloud or Alibaba Cloud), Cambricon must convince third-party data centers to adopt its Neuware software stack—an uphill battle against CUDA's decade-plus ecosystem moat.
Hygon: The x86-Compatible Dark Horse
Hygon (海光) occupies a unique position in China's AI chip landscape. As the only Chinese company with an x86 license (through its partnership with AMD), Hygon produces CPU+DCU (Deep Computing Unit) accelerator combinations that are architecturally similar to AMD's Instinct series.
The Hygon DCU K100 AI accelerator, based on a variant of AMD's CDNA architecture, provides roughly 100 TFLOPS of FP16 performance [6]. While that falls well short of the A100, let alone the H100, Hygon's strategic value lies in its compatibility with the x86 ecosystem and ROCm software stack.
For Chinese enterprises already running x86-based data centers, Hygon offers the lowest migration friction. Existing ROCm-compatible code can often be ported with minimal modification, and Hygon's CPU+DCU integrated platforms simplify deployment logistics.
Hygon's limitation is process technology. Subject to the same SMIC 7nm constraint as Huawei (and without Huawei's massive R&D budget), Hygon's chips cannot currently match TSMC-fabricated NVIDIA hardware on transistor density or power efficiency.
Performance Comparison: Chinese Chips vs. NVIDIA
| Specification | NVIDIA A100 | NVIDIA H100 | Huawei 910B | Cambricon 590 (est.) | Hygon K100 |
|---|---|---|---|---|---|
| Process Node | 7nm (TSMC) | 4nm (TSMC) | 7nm (SMIC) | 7nm (SMIC) | 7nm (SMIC) |
| FP16 (TFLOPS) | 312 | 989 | ~376 | ~300 | ~100 |
| Memory Bandwidth | 1,555 GB/s | 3,350 GB/s | ~1,200 GB/s | ~1,000 GB/s | ~800 GB/s |
| Interconnect | NVLink 600 GB/s | NVLink 900 GB/s | HCCS ~300 GB/s | Cambricon Link | PCIe 4.0 |
| Software Stack | CUDA | CUDA | CANN/MindSpore | Neuware | ROCm |
The performance gap is real but narrowing. On raw FP16 throughput, Huawei's 910B actually exceeds the A100. The critical differences lie in memory bandwidth (critical for LLM training), interconnect speed (critical for multi-chip scaling), and software ecosystem maturity.
The Ecosystem Gap: CUDA vs. CANN vs. ROCm
Hardware is only half the battle. NVIDIA's true moat is CUDA—the parallel computing platform that an estimated 4 million developers worldwide have built upon since 2007 [7]. Every major deep learning framework (PyTorch, TensorFlow, JAX) has CUDA-optimized backends as their default.
Huawei CANN (Compute Architecture for Neural Networks): CANN supports mainstream frameworks through adaptation layers and Huawei's own MindSpore framework. MindSpore has gained traction in Chinese academic and government research, but international adoption is negligible. The CANN toolkit has improved significantly—version 7.0+ offers near-feature-parity with CUDA for common inference workloads—but training large-scale LLMs still requires workarounds and custom optimization [8].
Cambricon Neuware: Cambricon's software stack supports PyTorch through a custom backend, but developer documentation and community support remain thin compared to CUDA. The installed user base is measured in hundreds, not millions.
Hygon/ROCm: Hygon benefits from AMD's open-source ROCm platform, which is architecturally compatible. However, Hygon's older CDNA-variant silicon doesn't support the latest ROCm 6.x features, creating a software-hardware mismatch.
The bottom line: Chinese chipmakers can match NVIDIA on transistor-level performance for specific workloads, but the ecosystem gap is measured in years, not months. Every researcher who writes `import torch` defaults to CUDA. Changing that default is the hardest problem in Chinese AI chip adoption.
Export Controls: Constraint and Catalyst
U.S. export controls have been a double-edged sword for China's AI chip industry. On one hand, they've cut off access to the most advanced NVIDIA hardware, forcing Chinese companies to buy domestic. On the other hand, they've also restricted access to the advanced manufacturing equipment (particularly ASML's EUV lithography systems) needed to produce cutting-edge chips.
The net effect has been a forced acceleration of domestic alternatives. Before 2022, Chinese cloud providers had little incentive to adopt domestic AI chips when A100s were readily available. By 2025, Baidu, Alibaba, and Tencent were all running significant domestic AI chip workloads—not by choice, but by necessity.
This "forced adoption" has created the one thing Chinese chipmakers needed most: real-world deployment feedback. Bugs that only surface at scale are being found and fixed. Software stacks are maturing under production load. Developer communities, while small, are growing organically.
2026–2030: The Catch-Up Timeline
2026: SMIC is expected to begin risk production on its N+3 (5nm-class) process using multi-patterning DUV. Huawei's Ascend 910C enters volume production. Cambricon's Siyuan 590 achieves broader commercial deployment. China's domestic AI chip market reaches an estimated $15 billion [9].
2027–2028: If SMIC's 5nm yields improve sufficiently, a new generation of Chinese AI chips could narrow the gap to NVIDIA's H100 series. The critical variable is whether domestic lithography alternatives (including SMEE's SSA/800-10W DUV scanner) can achieve production-grade reliability. Huawei's CANN ecosystem reaches critical mass with an estimated 500,000+ registered developers.
2029–2030: China's AI chip industry faces a fork in the road. If EUV alternatives (such as nanoimprint lithography or multi-beam direct-write) mature, China could potentially reach 3nm-class chips by 2030, narrowing the gap to whatever NVIDIA's architecture offers at that point. If not, China may remain locked at 5nm while TSMC and NVIDIA move to 2nm and beyond—widening the hardware gap even as the software ecosystem matures.
The consensus among industry analysts is that China will achieve "functional sufficiency"—the ability to train and deploy competitive AI models on domestic hardware—by 2027–2028, even if raw hardware performance lags NVIDIA's latest by 12–24 months [10].
Conclusion
China's AI chip trinity—Huawei, Cambricon, and Hygon—represents three different strategic approaches to the same problem. Huawei brings vertical integration and massive resources. Cambricon brings focused AI-first design expertise. Hygon brings x86 ecosystem compatibility. None of them can match NVIDIA today, but all three are improving faster than NVIDIA ever had to.
The export controls that were designed to hold China back may ultimately be the catalyst that creates NVIDIA's most serious competitors. The next four years will determine whether China's AI chip industry becomes a credible global alternative—or remains a domestically constrained workaround.
Frequently Asked Questions
1. Which Chinese AI chip is closest to NVIDIA in performance?
Huawei's Ascend 910B is currently the closest, delivering approximately 376 TFLOPS of FP16 performance versus the A100's 312 TFLOPS. However, the 910B trails the A100 in memory bandwidth and the H100 significantly across all metrics. On a price-performance basis for inference workloads, the 910B is competitive. For large-scale LLM training, NVIDIA still holds a clear advantage due to superior interconnect and software optimization.
2. Can Chinese AI chips run PyTorch and TensorFlow?
Yes, but with caveats. Huawei's CANN provides adaptation layers for PyTorch and TensorFlow, and Cambricon's Neuware offers similar compatibility. However, porting existing code often requires modifications—particularly for custom CUDA kernels, mixed-precision configurations, and distributed training setups. The porting effort ranges from hours (for standard model architectures) to weeks (for complex, CUDA-optimized training pipelines).
3. How do export controls affect Chinese AI chip development?
Export controls restrict both the chips China can import (A100, H100, H200, etc.) and the manufacturing equipment needed to produce advanced domestic chips (particularly EUV lithography systems). This creates a paradox: domestic demand for Chinese AI chips has surged (because NVIDIA hardware is unavailable), but domestic chips are constrained by manufacturing limitations (because advanced lithography equipment is also blocked). The net effect has been rapid ecosystem maturation but persistent hardware performance gaps.
4. Is SMIC capable of producing 5nm chips without EUV?
SMIC has demonstrated 7nm production using DUV multi-patterning (as seen in Huawei's Mate 60 Pro chipset). Extending DUV to 5nm is technically possible but comes with significantly lower yields and higher costs compared to EUV-based production. SMIC is expected to attempt 5nm risk production in 2026, but commercial viability depends on yield improvements that are not yet guaranteed.
5. What is the market share of Chinese AI chips in China?
As of 2025, domestic AI chips hold approximately 20–30% of the Chinese AI accelerator market, up from less than 5% in 2022 [9]. Huawei holds the largest share among domestic suppliers, followed by Cambricon and Hygon. The remaining 70–80% is dominated by pre-export-control NVIDIA inventory (A100, A800, H800) and smuggled hardware. As NVIDIA stockpiles deplete, domestic share is projected to reach 50%+ by 2028.
6. Will Chinese AI chips be available outside China?
Currently, no Chinese AI chip has meaningful international market share. Huawei's Ascend series is primarily deployed in Huawei Cloud within China. Cambricon and Hygon chips are used almost exclusively by Chinese enterprises and research institutions. Exporting Chinese AI chips would face both technical challenges (ecosystem adoption) and geopolitical barriers (potential U.S. sanctions on buyers of Chinese semiconductors). International availability is unlikely before 2028 at the earliest.
References & External Links
1. U.S. Bureau of Industry and Security — Export Controls on Advanced Computing and Semiconductor Manufacturing Items to China (September 2022): https://www.bis.doc.gov/index.php/documents/about-bis/newsroom/press-releases/3158-2022-10-07-bis-press-release-and-fact-sheet-1/file
2. Reuters — U.S. tightens rules on AI chip exports to China (October 2023): https://www.reuters.com/technology/us-tightens-rules-ai-chip-exports-china-2023-10-17/
3. TechInsights — Huawei Ascend 910B Die Analysis: https://www.techinsights.com/blog/huawei-ascend-910b-die-analysis
4. SemiAnalysis — China's Domestic AI Chip Landscape 2025: https://www.semianalysis.com/p/china-ai-chips
5. Cambricon Technologies — Official Product Documentation (思元系列): https://www.cambricon.com/
6. Hygon Information Technology — DCU Product Specifications: http://www.hygon.com/en/
7. NVIDIA Developer — CUDA Ecosystem Statistics: https://developer.nvidia.com/cuda-toolkit
8. Huawei Developer — CANN Documentation: https://www.hiascend.com/en/document/detail/en/CANNCommercial
9. IC Insights / Yole Group — China Semiconductor Market Forecast 2025–2030: https://www.yolegroup.com/strategy-insights/china-semiconductor-market/
10. McKinsey & Company — China's Semiconductor Industry: The Long March to Self-Sufficiency (2025): https://www.mckinsey.com/industries/semiconductors/our-insights/chinas-semiconductor-industry
*Article by electroniccomponent.com — your trusted source for electronic component insights, PCBA solutions, and semiconductor industry analysis.*
