Tencent's president, Martin Lau, announced during the Q1 2025 earnings call that the company has ample high-end GPUs thanks to prior acquisitions and effective training efficiencies. This shift in focus from scaling laws to utilizing smaller clusters has enabled Tencent to train large language models without needing additional silicon. By enhancing software optimization, Tencent aims to significantly improve inference efficiency, suggesting that these measures will prolong the lifespan of its existing GPU inventory while meeting future AI demands. The company is also moving towards training more application-specific models to maximize efficiency.
Tencent has a strong stockpile of chips and has developed efficient methods to train AI models, thus ensuring sufficient resources for years.
The shift in strategy moves away from the traditional belief of scaling laws and embraces smaller clusters for effective training, signaling an innovation in AI model training.
Lau stated that software optimization significantly enhances inference efficiency, proposing that enhancing this aspect could effectively double the capacity from existing GPUs.
Tencent's strategy includes investing in software improvements and training smaller models, focused on specific applications to ensure resource efficiency.
Collection
[
|
...
]