PolyU develops new GenAI training method that halves computing time while matching major models' performance

Hong Kong Polytechnic University's newly established Academy for Artificial Intelligence (PAAI) has developed a collaborative training approach that enables generative artificial intelligence models to achieve performance comparable to major systems while requiring significantly less computing power and training time.

The "collaborative generative AI" method shifts from traditional centralized training to a distributed approach, merging separately trained AI models into one large system. Researchers aim to open-source their "end-to-end FP8 low-bit training full solution" to enable broader participation in AI development.

Professor Yang Hongxia, executive director of PAAI, described current AI model training as "a game for the rich" due to massive computational resource requirements. "We hope through open-source technology to help those who couldn't build their own models to create customized systems at minimum cost," she said.

The PolyU method reduces video memory usage by approximately 24 percent and cuts training time by 48 percent compared to the mainstream BF16 approach, meaning similar results can be achieved with roughly half the computing power.

Whereas traditional foundation model training requires millions of GPU hours – estimated at 1-2 million GPU hours using 128 to 256 GPU cards training for half to one month for top global models – PolyU's InfiFusion model fusion technology needs only hundreds of hours to achieve similar or better performance.

Professor Yang compared the approach to "absorbing powers" where specialized small models excelling in areas like mathematics, coding or text understanding are merged to create comprehensive capabilities.

The team has already developed a medical foundation model and specialized "Cancer GenAI" for personalized treatment planning, currently collaborating with hospitals including Queen Elizabeth Hospital in Hong Kong and mainland institutions. Researchers are also exploring FP4 precision training to further reduce costs.

PolyU's Senior Vice President Christopher Chao emphasized that GenAI is transforming various fields and that AI represents the core of accelerating new quality productive forces. He stressed the importance of moving beyond theoretical research to practical applications that benefit industries.