Performance Comparison of Domestic GPUs: Who Is the King of Domestic Computing Power? - Zhihu

The latest discussion on Zhihu has sparked a lively debate about which domestic GPU truly reigns supreme in computing power. Chinese tech giants like Huawei, Biren Technology, and Moore Threads are pushing their chips to compete with global leaders like Nvidia, but the race is far from settled. Huawei’s Ascend 910B has shown impressive performance in AI training tasks, yet it still trails behind Nvidia’s H100 in raw benchmark scores. Meanwhile, Biren’s BR100 has captured attention for its strong performance in cloud gaming and graphics rendering, though software ecosystem gaps remain a challenge.

Moore Threads has emerged as a dark horse with its MTT S80, delivering solid results in video processing and basic AI workloads. However, the chip struggles with complex deep learning models that require high-precision floating-point calculations. The real battle lies not just in hardware specs, but in the software stack that makes these GPUs usable for developers. Chinese firms are racing to build compatible frameworks like PaddlePaddle and MindSpore, but they still lag behind Nvidia’s CUDA ecosystem in maturity and developer adoption.

Domestic GPUs have made huge strides in power efficiency, with some models achieving competitive performance per watt against Western alternatives. For instance, Huawei’s Ascend series now matches Nvidia’s A100 in certain inference tasks while consuming less energy. This efficiency is critical for China’s massive data centers, where electricity costs and cooling are major operational concerns. Yet, when it comes to raw compute for training large language models like DeepSeek, Nvidia still holds a clear lead in both speed and scalability.

The Zhihu analysis highlights that no single domestic GPU can claim the throne across all use cases. Huawei excels in AI training for government and enterprise clients, while Biren leads in graphics-intensive applications like virtual reality. Moore Threads is carving a niche in edge computing and lightweight AI tasks, where its low power draw is a big advantage. This fragmentation means that China’s tech ecosystem must rely on multiple chips for different workloads, unlike the unified dominance of Nvidia in the West.

Global discoveries in chip design, such as chiplets and advanced packaging, are now being adopted by Chinese firms to close the gap. For example, Biren’s BR100 uses a multi-chip module approach similar to Nvidia’s Hopper architecture, boosting performance without massive die sizes. These innovations are helping domestic GPUs reach parity in specific metrics, but the software ecosystem remains the biggest hurdle to widespread adoption.

In the end, the king of domestic computing power is not a single chip but the collective progress of China’s GPU industry. The Zhihu debate underscores that while no domestic GPU has dethroned Nvidia, the rapid pace of improvement signals a shifting landscape. For now, the real winner is the competition itself, driving down costs and spurring innovation that benefits the entire global tech community.