Archive — CN — technode
Ant Group cuts AI costs by 20% using Chinese chips
Ant Group has cut AI model training costs by 20% using Chinese-made chips from Alibaba and Huawei. By employing a method called mixture of experts (MoE), the company achieves results comparable to Nvidia's H800 chips, but continues using Nvidia and AMD hardware while aiming to reduce dependency on expensive GPUs, diverging from Nvidia's approach to AI advancements.
Key points
- Ant Group developed AI model training technology using Chinese chips.
- Cost reduction of 20% achieved with semiconductors from Alibaba and Huawei.
- The machine learning method, mixture of experts (MoE), enhances efficiency by splitting tasks.
- MoE models are gaining traction in the corporate AI landscape.
- Ant's training results are comparable to Nvidia's H800 chips.
- Ant continues to use Nvidia hardware but seeks to lessen reliance on high-end GPUs.
- Nvidia CEO Jensen Huang argues for the necessity of powerful chips for advanced AI, contradicting Ant's cost-cutting approach.
Why it matters
- This move aligns with the current trend of companies looking to localize AI technology and reduce supply chain risks, especially in the context of US-China tech tensions.
- The use of Chinese semiconductors reflects the burgeoning domestic AI chip industry amidst global semiconductor shortages.
- As companies push for cost efficiency in AI, strategies like MoE could shape how AI systems are developed and implemented by 2025.
Ant Group's initiative showcases a significant shift towards utilizing domestic technology in AI, promoting cost efficiency and innovation while challenging established narratives in the global tech landscape.