Huawei Ascend 910C Shows Promising Inference Capabilities Against Nvidia H100

In a significant development for China's domestic AI hardware sector, research conducted by DeepSeek suggests that Huawei's Ascend 910C processor is achieving approximately 60% of the inference performance benchmarked against Nvidia's formidable H100 GPU. This revelation, detailed in reports from various tech publications, underscores Huawei's ongoing efforts to advance its AI chip capabilities amidst stringent U.S. sanctions and limitations in accessing leading-edge semiconductor manufacturing technologies.

Ascend 910C: A Leap in Huawei's AI Hardware

The Ascend 910C represents an evolution of Huawei's earlier Ascend 910 AI training processor, originally introduced in 2019. While the Ascend 910's performance might be considered less competitive for the demanding tasks of large-scale AI training today, its successor, the Ascend 910C, demonstrates notable strength in inference operations. This new iteration is built using SMIC's second-generation 7nm-class process technology, known as N+2, a departure from the original Ascend 910's TSMC N7+ fabrication. The chiplet packaging design, a feature shared with its predecessor, houses around 53 billion transistors, indicating a sophisticated approach to chip architecture and integration.

Inference Performance: A Key Battleground

The comparison of the Ascend 910C against the Nvidia H100, particularly in inference tasks, highlights a critical area where Huawei is making strides. Inference, the process of using a trained AI model to make predictions or decisions, is a crucial component of deploying AI in real-world applications. Achieving 60% of the H100's inference performance is a substantial accomplishment, suggesting that Huawei's hardware is becoming a more viable option for companies seeking alternatives to Nvidia's dominant offerings. This is particularly relevant for the Chinese market, where reducing reliance on foreign technology is a strategic imperative.

Challenges in AI Training and Software Ecosystem

Despite the advancements in inference performance, experts, including those from DeepSeek, point out that AI training remains a domain where Nvidia holds a significant and largely undisputed lead. A primary reason cited for this is the deeply integrated and mature Nvidia software ecosystem, particularly its CUDA (Compute Unified Device Architecture) platform, which has been meticulously developed over two decades. This extensive software support provides a robust framework for the complex and iterative process of training large AI models. While inference performance can be optimized through specific hardware and software tuning, sustained and large-scale AI training workloads still present challenges for Chinese processors, requiring further enhancements in both hardware and software stacks.

The Role of Software Optimization and Future Architectures

The research from DeepSeek also emphasizes the pivotal role of software optimization in bridging performance gaps. DeepSeek's expertise in optimizing hardware and software, including tools that facilitate the conversion from CUDA to Huawei's CANN (Compute Architecture for Neural Networks), could significantly lessen the dependency on Nvidia. As AI models increasingly converge on architectures like the Transformer, which forms the basis for many large language models, the unique advantages of Nvidia's proprietary software ecosystem may become less pronounced. This shift could open up greater opportunities for alternative hardware solutions, such as Huawei's Ascend series, to gain market traction. The ability to optimize performance through algorithmic improvements and tailored software stacks, as demonstrated by DeepSeek, offers a pathway to more cost-effective AI solutions, especially for inference-intensive applications.

Navigating Sanctions and Technological Advancement

Huawei's progress with the Ascend 910C is particularly noteworthy given the U.S. government's sanctions, which restrict its access to advanced semiconductor manufacturing processes, such as those offered by TSMC. The fact that Huawei, in collaboration with foundries like SMIC, can produce competitive AI chips underscores the rapid advancement within China's domestic semiconductor industry. While challenges related to training stability and the comprehensive software ecosystem persist, the Ascend 910C's performance in inference marks a critical step forward. This development positions Huawei as a key player in the global AI hardware landscape, offering a tangible alternative and potentially reshaping the competitive dynamics in the AI chip market.

Looking Ahead: Competition and Collaboration

The ongoing advancements in AI hardware from companies like Huawei, coupled with the optimization efforts from research entities like DeepSeek, signal a dynamic and evolving market. While Nvidia continues to lead in many aspects of AI computing, the emergence of strong domestic alternatives in China is undeniable. The strategic importance of these developments extends beyond mere performance metrics; they are indicative of a broader trend towards technological self-sufficiency and the diversification of the global AI supply chain. The continued refinement of Huawei's Ascend chips and the development of supporting software infrastructure will be crucial in determining the extent to which they can truly challenge Nvidia's established dominance in the long term.

Tag: inference performance

Huawei Ascend 910C Emerges as a Strong Inference Contender, Challenging Nvidia