Huawei has taken the role of filling the void left by Nvidia’s H20, following the U.S. export restriction placed on chips supplied to China by introducing its next-generation Ascend 920 chip.
The Ascend 920 was unveiled shortly after U.S. export restrictions on Nvidia for H20 chips to China, which led to Nvidia losing ground in the Chinese AI market and giving room for other alternatives and competition to grow. Ascend 920 was built as a logical alternative for the Nvidia chip H20 and serves as a solid entry for this tech giant to gain footing in the AI market.
Huawei is a Chinese multinational corporation and technology company headquartered in Longgang, Shenzhen, China. It was founded in September 1987 by Ren Zhengfei. Before the company’s journey in the world of AI chips, Huawei was best known as one of the world’s three smartphone makers in 2018, alongside Apple and Samsung.
In May 2019, under Trump’s first administration, Huawei and several other companies were placed on the Entity List, a list that contains companies that are not allowed to do business with the U.S. The ban prevented Huawei from working with companies like Google, Qualcomm, and Intel, which served as a setback for its smartphone journey.
Huawei started its AI chip development journey following an announcement of the Ascend 310 chip in 2018 at its flagship event. It was officially launched in 2019.
Ascend 310 was designed for edge devices like security cameras, industrial sensors, and autonomous vehicles, consuming 8 W of power and delivering 8 TeraFLOPS (trillions of floating point operations per second) of half-precision floating point performance (FF16) and 16 TeraOPS (trillions of operations per second) in integer precision (INT8) calculation.
Ascend 910 was launched in August 2019, and described at the time as the world’s most powerful AI processor for data centers and cloud computing. It belonged to a series of Ascend-Max chipsets. Ascend 910 consumed 310 W of power, delivered 256 TeraFLOPS for half-precision floating point (FP16) operations, and 512 TeraOPS in integer precision calculations.
Ascend 910 was developed alongside MindSpore, a computing framework for supporting development for AI applications.
After its U.S. sanctions, Huawei continued to research and develop ways to improve AI performance, integrate with Huawei Cloud AI services, and develop an in-house AI framework to reduce reliance on the U.S. for chips.
In 2022, Huawei launched Ascend 910B, its totally homegrown AI chip, which was set to rival Nvidia’s A100. It had a computing power of 400 TeraFLOPS for half-precision floating point (FP16) and 800 TeraOPS in integer precision calculations. It was manufactured using SMIC’s second-generation 7nm-class process technology.
In October 2024, Ascend 910C was launched and is expected to be mass-produced in Q1/early Q2 of 2025. The 910C is an evolution of the 910B model and offers a computing power equal to 60% of Nvidia’s H100. It has a transistor count of almost 53 billion, a 7nm SMIC manufacturing process called N+2, a computing power of 800 TeraFLOPS for FP16, a memory bandwidth of 3.2 TB/s, and consumes 310 watts of power.
The 910C was designed for workloads involving natural language processing, predictive analysis, and computer vision. To put the chip on par with the H100, Huawei compensates by deploying the chips in large clusters called the CloudMatrix 384. It is said, however, that the CloudMatrix 384 outperforms Nvidia’s GB200 NVL72 and consumes 3.9x the power required for the GB200.
The Ascend 920 was announced a day after the U.S. government implemented export restrictions on AI chips, particularly Nvidia’s H20.
The chip is built using SMIC’s 6nm node (which makes it smaller than its predecessors); it delivers over 900 TeraFLOPS per BF16 (brain floating point), a 4 TB/s bandwidth, and utilizes HBM3 modules. Ascend 920 is designed to enhance scalability and efficiency. It is anticipated to be mass-produced in the second half of 2025.
