Huawei vượt qua lệnh trừng phạt của Mỹ bằng chiến lược sản xuất 700.000 chip AI
-
Huawei đã vươn lên mạnh mẽ sau lệnh trừng phạt của Mỹ năm 2019, với dự kiến xuất xưởng 700.000 bộ xử lý AI Ascend trong năm 2025 (theo Mizuho), gồm các dòng 910A, 910B và 910C.
-
Dòng chip Ascend 910C mới có kiến trúc hai khuôn (dual-die), kết nối tốc độ cao, bộ nhớ tích hợp 8 tầng – hỗ trợ xử lý AI hiệu quả hơn đáng kể.
-
Huawei phát triển hệ thống CloudMatrix384 dùng 384 chip Ascend và 192 CPU Kunpeng, có khả năng xử lý 1.943 tokens/giây mỗi NPU và thời gian phản hồi dưới 50 mili giây/tokens trong giai đoạn giải mã AI.
-
CloudMatrix vượt qua framework phục vụ LLM của Nvidia là SGLang trong thử nghiệm mô hình DeepSeek R1 – cho thấy Huawei đang dẫn đầu về quy mô triển khai AI lớn.
-
Báo cáo từ SemiAnalysis cho rằng tuy chip Huawei tụt hậu một thế hệ so với Mỹ, nhưng hệ thống tổng thể lại vượt trước Nvidia và AMD về thiết kế và hiệu năng cụm.
-
Nvidia đã chịu tổn thất 4,5 tỷ USD vì chip H20 bị hạn chế xuất sang Trung Quốc, dự kiến mất thêm 8 tỷ USD doanh thu trong quý tới.
-
Các công ty AI lớn như iFlytek và SenseTime đã chuyển sang dùng chip nội địa như Ascend để tránh rủi ro từ lệnh cấm.
-
Huawei nhận được hỗ trợ từ chuỗi cung ứng nội địa, trong đó SiCarrier trình làng thiết bị sản xuất chip mới tại Semicon China 2025.
-
Trong khi Mỹ hạn chế phần mềm thiết kế EDA, các công ty Trung Quốc như Empyrean, Primarius và Semitronix đang nổi lên với khả năng cung cấp 80% công cụ EDA nội địa.
-
Ren Zhengfei, CEO Huawei, khẳng định không lo lắng về khó khăn từ lệnh cấm và tập trung vào tiến từng bước trong chiến lược tự chủ công nghệ.
-
Huawei được cho là đã xây dựng một hệ sinh thái phần cứng AI quy mô lớn, dẫn đầu nỗ lực tự lực về bán dẫn tại Trung Quốc.
📌 Huawei bất chấp lệnh cấm Mỹ, kỳ vọng xuất xưởng 700.000 chip AI Ascend trong năm 2025 với hệ thống CloudMatrix vượt hiệu suất của Nvidia tại Trung Quốc. Hệ thống dùng chip 910C xử lý 1.943 tokens/giây, dưới 50 mili giây/tokens. Huawei hiện đóng vai trò then chốt trong chiến lược AI và tự chủ bán dẫn quốc gia, với hệ sinh thái phần cứng đang phát triển mạnh.
https://www.scmp.com/tech/big-tech/article/3315239/how-huaweis-silicon-strategy-defies-us-sanctions-advance-chinas-ai-ambitions
How Huawei’s silicon strategy defies US sanctions to advance China’s AI ambitions
Huawei is expected to ship about 700,000 Ascend AI processors in 2025, according to Mizuho estimates
Ann Caoin ShanghaiandWency Chenin Shanghai
About three months after Huawei Technologies was added to Washington’s trade blacklist in 2019, the company’s then-rotating chairman, Eric Xu Zhijun, unveiled what he proclaimed as the “world’s most powerful AI processor” – the Ascend 910 – designed for training artificial intelligence (AI) models.
Huawei’s advanced AI chip initiative, however, suddenly faced a major obstacle a year later in August 2020, when the US Commerce Department tightened restrictions by barring the sale of semiconductor products and services – sourced from anywhere with US technology – to the company and its affiliates without a requisite licence.
As a result, Huawei supplier Taiwan Semiconductor Manufacturing Co, the world’s largest and most advanced contract chipmaker, ceased doing business with the Chinese firm and its integrated circuit (IC) design unit HiSilicon to comply with US curbs.
At the time, the prognosis appeared dire for Huawei, according to some analysts. “If enough companies comply globally, Huawei’s ability to generate workarounds will be severely undercut, putting its continued existence as a viable commercial entity in doubt,” said Paul Trolio of New York-based political risk consultancy Eurasia Group.
Fast-forward to 2025, and Huawei has remained resilient in the face of US sanctions.
Jensen Huang, founder and CEO of AI chip giant Nvidia, has been the most prominent industry leader to recognise the resurgence of Huawei in the IC sector.
“All in all, the export controls were a failure. The facts would suggest it,” Huang told reporters on the sidelines of last month’s Computex expo in Taipei. He called on the White House to lower barriers to AI chip sales before American firms cede the China market to rivals like Huawei.
Huang’s assessment followed Nvidia’s disclosure in April that the US government now required a licence to export its H20 chips to China. The company incurred a US$4.5 billion charge in its first quarter ended April 27, owing to excess inventory and purchase obligations for H20.
For the second quarter, Nvidia expects a US$8 billion revenue loss from the H20 chips ban. The H20 graphics processing unit (GPU) was designed for the China market after earlier US export controls.
The performance of Ascend chips against Nvidia’s in-demand GPUs was put under the spotlight this week, following the release of a technical paper that was jointly written by researchers from Huawei and Chinese AI infrastructure start-up SiliconFlow.
According to the paper, Huawei’s Ascend-powered advanced data centre architecture – CloudMatrix 384, along with the serving solution CloudMatrix-Infer – outperformed the Nvidia GPU-based SGLang fast-serving framework for large language models (LLMs), on both the inference and decoding phases, in running DeepSeek’s R1 reasoning model.
The CloudMatrix system, which the paper projected to “reshape the foundation of AI infrastructure”, consisted of 384 Ascend 910C neural processing units (NPUs) and 192 Kunpeng server central processing units, interconnected through a unified bus providing ultra-high bandwidth and low latency.
The highly efficient architecture reflects Huawei’s commitment to overcoming US tech restrictions, as the company pushes the boundaries of AI system performance.
According to an earlier assessment by SemiAnalysis, Huawei’s CloudMatrix directly competes with Nvidia’s GB200 NVL72 rack-scale system in alleviating bottlenecks for data centres.
In the decode phase of generating output from an AI model, the Huawei-SiliconFlow paper’s findings showed that CloudMatrix recorded 1,943 tokens per second per NPU for a 4,000-length key-value cache – a memory structure that enables more efficient use of AI processors.
The same phase showed output generation times consistently below 50 milliseconds per token, yielding an efficiency of 1.29 tokens per second per trillion floating-point operations per second.
“These results collectively establish CloudMatrix384, in combination with our peer-to-peer serving solution CloudMatrix-Infer, as a scalable, high-throughput and production-grade platform for large-scale LLM deployment,” the paper said.
It also provided details of Huawei’s 910C AI processor for the first time. The chip has a dual-die architecture, integrating two identical computing dies in a single package. These dies share eight stacks of on-package memory and communicate via a high-bandwidth connection fabric, which enables rapid data transfer that is crucial for intensive AI workloads.
“Huawei is a generation behind in chips, but its scale-up solution is arguably a generation ahead of Nvidia and AMD’s [Advanced Micro Devices] current products on the market,” the SemiAnalysis report said.
That echoed remarks made by Huawei founder and CEO Ren Zhengfei in a recent front-page interview published by the People’s Daily, the official mouthpiece of the Communist Party. Ren acknowledged that the Shenzhen-based firm’s Ascend chips still lagged behind those of US competitors by “a generation”.
He added, however, that using methods such as “stacking and clustering” resulted in computing performance comparable to the most advanced systems in the world.
“A single chip’s performance is not as important as the combined power of a group of chips working together in parallel,” said tech investor Kevin Xu, the founder and chief information officer of Interconnected Capital.
He pointed out that building large-scale systems requires networking expertise – a proficiency of Huawei, which remains the world’s largest vendor of telecommunications equipment.
“US export control has worked in slowing AI development progress in China, but it has also accelerated China’s push to indigenise every step of advanced AI chipmaking,” Xu said. Nvidia is no longer a reliable source of AI chips to the mainland, even if the company is allowed to sell in the market again, he added.
Meanwhile, the Trump administration’s cryptocurrency and AI tsar David Sacks warned that China had grown adept at evading US export controls and was, at most, two years behind American semiconductor design capabilities, according to a Bloomberg report on Friday.
He pointed out that DeepSeek’s breakthrough AI models earlier this year showed how China could still advance even with export controls in place.
Sacks said tight US restrictions on sales of AI chips to American allies – based on the Biden-era “AI Diffusion” rule, which the Trump administration has rescinded – could have unwittingly created an opening globally for Huawei and other Chinese companies.
“If we are too restrictive in terms of US sales to the world, I think that there will be a time when we kick ourselves and say, ‘All of a sudden Huawei is everywhere when we used to have the market to ourselves. Why didn’t we take advantage of that and lock it in?’” he said.
Those remarks echo comments made by Nvidia’s Huang on the sidelines of last week’s VivaTech conference in Paris, where he warned that Huawei was in a position to expand its semiconductor business should US chip export curbs stay in place.
“If the United States doesn’t want to participate in China, Huawei has got China covered,” Huang said. “Huawei [also] has got everybody else covered.”
Still, Huawei was expected to ship no more than 200,000 of its advanced AI processors this year, US Under Secretary of Commerce for Industry and Security Jeffrey Kessler told a congressional hearing last week. He estimated that “most or all of” those chips would be supplied to enterprises within China.
That figure would pale in comparison to the more than 1 million China-specific H20 GPUs produced by Nvidia in the last nine months of 2024, according to a January report from US research firm SemiAnalysis.
Estimates by analysts from Mizuho Securities, however, found Huawei could ship more than the number Kessler told US lawmakers.
About 700,000 units of Huawei’s Ascend 910 series – including the 910A, 910B and 910C – were expected to ship in 2025, despite yield challenges faced by the mainland’s biggest contract chipmaker, Semiconductor Manufacturing International Corp, at the 7-nanometre (nm) node, according to Mizuho.
An aerial view of the factory of Shanghai-based Semiconductor Manufacturing International Corp in Shenzhen, southern Guangdong province. Photo: VCG via Getty Images.
Speculation on how Huawei has managed to pull off increased production of Ascend chips has pointed to the mobilisation of China’s chip equipment supply chain.
SiCarrier, a Chinese semiconductor equipment maker with ties to Huawei, gained wide acclaim for its first public display of dozens of new chip manufacturing and testing machines at Semicon China in March. That raised conjecture about SiCarrier’s possible contribution to the local development of Huawei’s 7-nm chip, initially used in the Mate 60 Pro 5G handset released in 2023.
A steady output of Ascend chips would augur well for Chinese AI companies, such as iFlytek and SenseTime, that have already shifted to buying locally made chips.
Liu Qingfeng, founder and chairman of voice-recognition specialist iFlytek, said earlier this month that his firm used domestic chips, such as Huawei’s Ascend 910B, even though that could delay AI model development by three months. The company has been barred from buying advanced US chips since it was added to Washington’s trade blacklist in October 2019.
Hong Kong-founded SenseTime, meanwhile, has been investing heavily in building up its AI infrastructure with Chinese-made chips – including those from Huawei, Cambricon Technologies and Hygon Information Technology – to mitigate risks from the ongoing US-China tech war.
Yang Fan, co-founder of SenseTime and president of its AI infrastructure unit SenseCore, said in April that the number of domestic chips in the firm’s data centres was “growing very fast” and “continues to rise every year”.
Still, the immediate challenge for domestic chip suppliers is meeting the potentially large demand from China’s biggest tech companies, led by Alibaba Group Holding and Tencent Holdings, once they have used up existing inventories of Nvidia chips. Alibaba owns the South China Morning Post.
“Big Tech companies have stepped up their capital expenditure over the past quarters and piled up their inventory,” said Alex Yao, head of China equity research at JPMorgan. “That probably means for the next six to twelve months, they don’t need to worry about computing power.”
In another segment of China’s chip supply chain, Empyrean Technology, Primarius Technologies and Semitronix are getting increased attention as alternative suppliers of electronic design automation (EDA) software after Washington directed the world’s leading vendors to halt sales on the mainland under new US export restrictions.
Top EDA suppliers Cadence Design Systems, Synopsys and Siemens EDA have confirmed that they received notices from the US Commerce Department’s Bureau of Industry and Security about the restrictions on their software, which could be used to develop advanced AI chips.
Empyrean sees both challenges and opportunities in the situation. The firm is currently able to supply around 80 per cent of the total 58 EDA tools used in the entire chipmaking process, said Yu Han, a senior marketing director at Empyrean, at an industry event on Friday.
Yu said Empyrean aims to supply all of those tools eventually, transforming the company into one of the world’s top-tier EDA suppliers. He pointed out that the most difficult part was establishing a solid ecosystem, which requires “collaboration with China’s semiconductor industry”.
Apart from lingering doubts about the efficiency of home-grown processors, users of these chips could also face compliance issues once the AI models they developed were adopted in overseas markets, according to a recent analysis from Shanghai-based semiconductor consultancy ICWise.
JPMorgan’s Yao said Chinese AI model developers put greater focus on the efficiency of their hardware deployment.
But with US tech curbs still in place, Huawei looks poised to benefit as the company continues to raise the bar on AI chip performance using all the techniques at its disposal, while leading China’s efforts in semiconductor self-sufficiency.
When asked about the US sanctions in his People’s Daily interview, Huawei’s Ren said he had rarely given that a thought because it was “useless” to ponder over difficulties.
“I don’t think about difficulties,” he said. “I just move forward one step at a time.”
Ren was not available for an interview, and Huawei’s spokesman declined to comment.
Thảo luận
Follow Us
Tin phổ biến
TAG
AI giáo dục
AI sinh-y-duoc
AI nghệ thuật
AI pháp lý-quản trị-chủ quyền
AI models
AI xã hội
AI prompts
AI kiến thức-khóa học
AI công nghiệp-lĩnh vực
AI edge
AI viễn thông
AI tools
AI chính phủ
AI cybersecurity
AI so sánh
AI đạo đức
AI tips
AI market
AI quân sự
AI an toàn-an ninh-techwar
AI việc làm
AI doanh nghiệp
OpenAI ChatGPT
AI môi trường-năng lượng
AI skill-talent
AI & công nghệ khác
AI nghiên cứu
AI chips-hardware-compute
AI vs con người
AI coding assistant
AI mở-nguồn mở
AI năng suất
AI startup-M&A
AI tương lai
AI báo chí
AI data
AI bản quyền
AI PC
AI riêng tư
AI deepfake-ảo giác-ANTT
AI ảnh-video-music-âm thanh
AI minh bạch
AI nhỏ
AI nông nghiệp-thực phẩm
AI ngân hàng-tài chính
AI giao thông
AI smartphone
AI robotics-auto-agents
AI consumer devices
AI manufacturing
AI benchmark
Telecom
AI thành công-thất bại
Digital
Semi-Cloud-DC-Green
HTS
STI
FAQ