Google introduced AlphaChip, a reinforcement learning method for artificial intelligence for chip design. It promises to significantly speed up chip design, as well as improve chips in terms of performance, power and area. Google used this solution when designing TPU (Tensor Processing Unit) AI accelerators, and other companies, including MediaTek, also used it.
Image source: Google
The chip design or chip design is the longest and most labor-intensive stage in the development of a semiconductor component. Synopsys, which makes chip design software, has implemented AI to solve this problem, but its product comes at a very high price. Google decided to democratize this approach. Currently, creating a die plan for a complex chip such as a GPU takes about two years if it is designed by humans. Less complex components can be designed in a few months, but the process involves millions in costs because large manufacturers employ a significant number of specialists. AlphaChip, according to Google, solves the problem in a few hours. Moreover, this system delivers superior results, optimized for performance and energy efficiency. Google also showed off a graph that shows the reduction in wire length in previous TPU versions and the new Trillium.
AlphaChip is based on a reinforcement learning model, in which the AI performs an operation in a predetermined environment, examines the results, and learns from the experience to improve its performance in the future. In the case of AlphaChip, the AI treats chip plan design as a kind of game, with one circuit component placed on the board per turn. A neural network helps build a graph of relationships between components, and the more layouts the system creates, the higher the quality of its work.
Google has been using AlphaChip to develop TPU AI accelerators since 2020, where the company runs large-scale AI models and cloud services. Transformer models run on these processors – this architecture is used in Gemini and Imagen. The AlphaChip system has helped improve the design of each successive generation of TPUs, including the latest Trillium, reducing development time and delivering higher performance. However, both Google and MediaTek use this system for a limited set of blocks, and a significant part of the work is still done by humans.
In addition to Google TPU, AlphaChip was used in the design of MediaTek Dimensity mobile 5G chips, which are widely used in current smartphones. The system was pre-trained on a wide range of chips, Google says, allowing it to generate increasingly efficient layouts as time goes on. Humans learn quickly, and AI learns even faster.
AlphaChip’s success has encouraged Google to continue introducing AI into various stages of chip design, including logic synthesis, macro selection and timing optimization – something Synopsys and Cadence offer for big money. The company believes that in the future, AlphaChip can be used throughout the entire chip development cycle from architectural design to layout and production – optimization using AI will help speed up chips, make them more compact, energy efficient and cheaper. In the future, the solution will be used not only for Google server accelerators and MediaTek mobile platforms. AlphaChip development continues.