Huawei's Ascend 910C Challenges NVIDIA's AI Dominance, Achieving 60% of H100 Inference Power

The AI chip landscape is heating up as Huawei's Ascend 910C emerges as a competitor to NVIDIA's H100. Recent reports suggest that the Ascend 910C, a modified version of the 2019 Ascend 910, achieves an impressive 60% of the NVIDIA H100's inference performance. This development signifies China's growing capabilities in the AI sector, particularly amidst ongoing US restrictions.

DeepSeek's Revelation: A Boost for Chinese AI?

Interestingly, these performance figures come from researchers at DeepSeek, known for their efficient AI models that challenge NVIDIA's dominance. DeepSeek's testing of Huawei's chips highlights China's progress in AI chip development, aiming to reduce reliance on foreign technology. This could be a significant step towards AI independence for China.

Ascend 910C: A Closer Look at Huawei's AI Chip

The Huawei Ascend 910C is reportedly a complete in-house design, utilizing SMIC's (Semiconductor Manufacturing International Corporation) 7nm N+2 process and incorporating 53 billion transistors. Similar to its predecessor, it employs chiplet packaging. However, unlike the original Ascend 910 manufactured by TSMC, the 910C's compute chiplet is produced locally by SMIC. This further strengthens China's domestic semiconductor capabilities.

Inference vs. Training: NVIDIA Still Leads in AI Training

While the Ascend 910C demonstrates strong inference capabilities, reports indicate that NVIDIA remains the leader in AI training. Inference refers to using existing AI models to make predictions, while training involves creating and refining these models.

Yuchen Jin from DeepSeek noted that Chinese chips face challenges in long-term training reliability. NVIDIA's well-established CUDA ecosystem, developed over two decades, offers a deeply integrated hardware and software solution that is difficult to match. While inference can be optimized, Huawei needs to further enhance both its hardware and software to handle demanding and sustained training workloads.

• Inference: Using existing AI models for predictions.
• Training: Creating and refining AI models.

Challenges for NVIDIA: DeepSeek Investigation and Potential Tariffs

NVIDIA faces challenges beyond competition from Huawei. Investigations are underway regarding whether DeepSeek illegally acquired NVIDIA chips through third-party buyers. In addition, potential tariffs on chips could further disrupt the market.

Related Articles

• Huawei Rumored to Launch New High-End AI Chip, Potentially Rivaling NVIDIA’s H100
• Huawei Reportedly Begins Sampling of Ascend 910C to Rival NVIDIA

The Road Ahead: Competition and Innovation in the AI Chip Market

Huawei's Ascend 910C represents a significant step forward for China's AI industry. While NVIDIA currently maintains a lead in AI training, the progress of Huawei and other Chinese companies is driving competition and innovation. The future of the AI chip market will likely be shaped by continued advancements in both hardware and software, as companies strive to overcome limitations and build comprehensive AI solutions. This competition could bring better performance and more options in the AI server market using technologies such as High Bandwidth Memory (HBM).