Archive — US — turtlesai
Huawei challenges Nvidia: the cluster cloudmatrix exceeds expectations
Huawei has introduced the Cloudmatrix 384 AI cluster, outperforming Nvidia's GB200 NVL72 system. Featuring 384 Ascend 910C chips, it delivers 300 petaflops, surpassing Nvidia’s 180 petaflops. This move reflects Huawei's resilience in the AI space despite U.S. restrictions while the CEO of Nvidia acknowledges Huawei's advancements.
Key points
- Huawei's Cloudmatrix 384 integrates 384 Ascend 910C chips, outperforming Nvidia's GB200 NVL72 system.
- Ascend 910C achieves 60% of the NVIDIA H100 inference performance as per Deepseek's tests.
- Production yield of Ascend 910C chips has improved from 20% to 40%, targeting 60% by 2025.
- Huawei's cann software reduces dependence on Nvidia's CUDA ecosystem, compatible with various AI frameworks.
- The Cloudmatrix 384 aims to meet increasing domestic demand and bolster Huawei's market position.
Why it matters
- Huawei's advancements in AI technology align with the broader trend of increasing technological self-sufficiency amidst international sanctions.
- This development reflects a significant shift in the competition between U.S. and Chinese tech companies, especially in the AI sector.
- By 2025, the demand for AI infrastructures is expected to grow, emphasizing the importance of domestic production capabilities.
Huawei's launch of the Cloudmatrix 384 cluster showcases its growing capabilities in AI and challenges Nvidia’s dominance, highlighting a pivotal shift in the global tech landscape due to geopolitical tensions.