Google's innovative approach of using artificial intelligence to design AI chips is revolutionizing the semiconductor industry. The key to their AlphaChip system lies in its reinforcement learning method, which treats chip layout as a game.
AlphaChip operates by placing one circuit component at a time on a blank grid until all components are positioned. The system is then rewarded based on the quality of the final layout, with optimized performance being the ultimate goal or "win" condition of the game.
To train the AI, Google initiated ten thousand simultaneous "games," allowing AlphaChip to practice designing layouts on ten thousand chips while collecting data and continuously improving its optimization strategies.
One of AlphaChip's most impressive features is its ability to learn the relationships between interconnected chip components and generalize across different chip designs, enabling it to improve with each layout it creates.
To design Tensor Processing Unit (TPU) layouts, AlphaChip first practices on various chip blocks from previous generations, including on-chip and inter-chip network blocks, memory controllers, and data transport buffers. This process allows AlphaChip to gain valuable experience before Google deploys it on current TPU blocks to generate high-quality layouts.
The results have been remarkable, with AlphaChip demonstrating superior performance in chip area utilization, power efficiency, and wirelength compared to human experts. Perhaps most impressively, the AI can generate superhuman or comparable chip layouts in under 6 hours, a task that typically takes human experts weeks to complete under similar circumstances.
AlphaChip's success has led to an increased reliance on TPUs at Google. In December 2023, Google released three different versions of its Gemini AI model, which required a large number of Cloud TPU v5p chips for training. More recently, in May 2024, Google unveiled its 6G TPU chip, Trillium, which offers enhanced functionality for matching training models.
AlphaChip has played a crucial role in improving the design of each TPU generation, with the number of chip blocks it can optimize increasing from 10 to 25. Furthermore, TPU chips are gaining broader market recognition beyond Google's internal use. In July 2024, Apple successfully trained AI models using Google's Cloud TPU clusters. Apple also noted that over 60 percent of Generative AI (GenAI) startups and 90 percent of GenAI unicorns utilize Google Cloud's AI infrastructure and Cloud TPU service.
Article translated by Charlene Chen
