Key Insights
- Performance: The Ascend 920 offers approximately 90% of the Nvidia H100’s BF16 compute performance, making it a strong contender in AI training and inference tasks.
- Memory Bandwidth: With 4 TB/s, the Ascend 920 surpasses the H100’s 3.35 TB/s, potentially benefiting memory-intensive AI workloads.
- Manufacturing Process: Huawei’s reliance on SMIC’s 6nm process, which lacks EUV lithography, may impact yield and efficiency compared to Nvidia’s TSMC 5nm process.
- Power Efficiency: While the Ascend 920 shows improvements over its predecessor, it still lags behind Nvidia’s H100 in terms of power efficiency.
- Software Ecosystem: Nvidia’s CUDA and AI software stack are industry standards, whereas Huawei’s ecosystem is still maturing, which may affect adoption outside China.
Conclusion
The Huawei Ascend 920 represents a significant step forward for China’s AI chip capabilities, offering competitive performance metrics that approach those of Nvidia’s H100. However, challenges remain in manufacturing efficiency, power consumption, and software ecosystem maturity. For domestic applications within China, especially given export restrictions on Nvidia’s H20, the Ascend 920 is poised to be a viable alternative. Globally, Nvidia’s H100 maintains advantages in efficiency, software support, and availability.
Note: The Ascend 920 is expected to enter mass production in the second half of 2025
