Huawei Tests Powerful AI Chips in China Amid Nvidia, US Ban for Ascend 910C
Despite the US restrictions on Nvidia, Chinese firms have found ways to obtain the company's chips, including Nvidia's advanced H100 GPUs. One such company, Huawei, has developed its own chips to compete with Nvidia's offerings, specifically the Ascend 910C.
The Ascend 910C is a successor to the Ascend 910B and is expected to be shipped to clients starting in October 2024. According to reports, the Ascend 910C can rival Nvidia's H100 chips, delivering about 60% of the H100's single-chip AI compute power. Huawei achieves this by combining two Ascend 910B processors.
| Feature / Capability | Huawei Ascend 910C | Nvidia H100 | |--------------------|---------------------------------------|------------------------------------------| | Architecture & Process | Combination of two Ascend 910B chips; reportedly based on SMIC's 7nm N+2 process for successor 910D[3] | Built on advanced 4nm node technology; latest generation GPU architecture[3] | | Performance (single chip) | About 60% of Nvidia H100 performance in AI tasks[1][5] | Industry-leading AI training performance | | Aggregate System Scale | CloudMatrix 384 system with 384 Ascend 910C chips delivers ~300 petaFLOPS BF16 compute, nearly double Nvidia GB200 NVL72 system[2][3] | GB200 NVL72 system delivers ~138 petaFLOPS BF16 compute | | Memory | CloudMatrix 384 boasts 3.6× greater total memory capacity and 2.1× higher memory bandwidth than GB200 NVL72 system[2] | Uses HBM3 memory with about 3 TB/s bandwidth[3] | | Energy Efficiency | Ascend 910D successor chip offers 12% better energy efficiency than H100; CloudMatrix system consumes over 4× the power of Nvidia’s system at full load[3][2] | Highly energy-efficient 4nm H100 with tensor cores for acceleration | | Software Ecosystem | Huawei's MindSpore AI framework is growing but lacks CUDA’s maturity and extensive adoption globally[3][4] | CUDA platform with ~90% global adoption, stable and mature software stack[3][4] | | Interconnection | All-to-all topology connecting 384 chips to maximize parallel performance at system scale[2] | NVLink and other high-speed interconnects for multi-GPU scaling |
Huawei's approach emphasizes scale and integration, leveraging quantity, memory capacity, and high interconnectivity over individual chip performance. This strategy results in superior total AI compute throughput in some configurations, but it comes with higher energy consumption and a less mature software ecosystem, which may hinder adoption outside China’s infrastructure-focused environment.
Meanwhile, Nvidia maintains a clear edge in chip architecture, power efficiency, and global software ecosystem dominance. However, the demand for high-performance chips is rising in China, and there is a big appetite for advanced semiconductors in the market. The Ascend 910C chips have been used to train the vast majority of top-performing large language models in China, and they are available in Chinese online markets at a significantly high value.
Huawei's move is considered a threat to Nvidia, as the Ascend 910C chips are intended for Chinese companies looking to shift away from Nvidia chips. In response, Nvidia is expected to deliver over one million H20 GPUs in China this year. The Ascend series of chips from Huawei has been popular with Chinese firms, and the success of the Ascend 910C could signal a shift in the chip market within China.
- Huawei's Ascend 910C AI chip can rival Nvidia's H100 - South China Morning Post
- Huawei Ascend 910C AI Chip Offers Superior Compute Power to Nvidia's H100 - Tom's Hardware
- Huawei Ascend 910C AI Chip: Architecture, Performance, and Specs - AnandTech
- Huawei Ascend 910C vs Nvidia H100: A Comparison - TechRadar
- By matching nearly 60% of Nvidia's H100's single-chip AI compute power, the Huawei Ascend 910C showcases significant rivalry within the realm of science and technology.
- The Huawei Ascend 910C, combating Nvidia's H100, shines not only in performance but also in scale and integration, offering a threat to Nvidia in the growing Chinese chip market.
