Key Takeaways
1. Huawei plans to introduce AI infrastructure technology to improve management of various chips, including its Ascend series and Nvidia’s chips.
2. The new software solution aims to increase AI chip utilization from 35% to 70%, effectively doubling the efficiency of AI data center clusters.
3. Huawei is competing with Nvidia and other Western companies for AI computing power, focusing on quantity to offset quality due to restrictions on high-performance chips.
4. The strategy of commoditizing AI resources emphasizes the need for power to support numerous data centers instead of just focusing on individual chip capabilities.
5. Huawei’s upcoming announcement at the AI Container Application Forum may showcase how software can enhance performance despite hardware limitations.
Huawei is expected to reveal a sophisticated AI infrastructure technology that aims to streamline the management of various chips, including its own Ascend series and those from Nvidia.
Boosting AI Chip Efficiency
This software-driven solution is projected to elevate the utilization rate of AI chips from the current average of 35% to 70%, effectively doubling the efficiency of the AI data center clusters. By masking the differences in hardware, this approach enhances resource allocation for AI training and inference tasks.
Competing on a Global Scale
As the leading AI chip developer in China, Huawei is at the center of the ongoing battle for AI computing power supremacy against Nvidia and other significant Western GPU companies. While it may be difficult to match Nvidia’s cutting-edge Blackwell AI chip architecture with existing production capabilities in China, Huawei is pursuing strategies that focus on increasing quantity to compensate for quality.
Due to restrictions on acquiring high-performance chips from Nvidia, which are both pricey and politically sensitive, China is making efforts to commoditize AI computing resources. Huawei has been grouping its numerous lower-end Ascend GPUs to operate open-source AI models, like DeepSeek, which require significantly less computing power compared to ChatGPT or Google’s Gemini, yet still manage to deliver similar performance levels.
A Shift in AI Strategy
This strategy of commoditizing AI appears to be effective currently, as it shifts the competition towards the power needed to support numerous AI data centers, rather than solely on chip capabilities or individual large language models (LLMs). For example, TikTok’s parent company ByteDance is leveraging the most popular chatbot in China, which also happens to be the largest consumer of AI computing power. Its daily demand has surged from 4 trillion tokens last year to over 30 trillion tokens now, closely rivaling Google’s consumption of 43.2 trillion tokens per day.
The upcoming announcement of Huawei’s integrated AI infrastructure control at the 2025 AI Container Application Implementation and Development Forum on November 21 could further exemplify China’s tactic of “using software enhancements to compensate for inferior hardware.”
It remains uncertain how Huawei aims to achieve a doubling of the AI chip optimization rate through infrastructure control improvements that can harmonize resources across different types of GPUs, such as its Ascend chips, Nvidia’s Blackwell, and those from other manufacturers, to boost the overall efficiency of computing clusters.
Source:
Link