Google announces chip design AI 'AlphaChip', already used to design smartphone chips and AI-specific chips
Google DeepMind, Google's AI research division, has announced the AI ' AlphaChip ' that can design chip layouts. AlphaChip has already been introduced into actual chip development and is being used in chip designs by Google and MediaTek.
How AlphaChip transformed computer chip design - Google DeepMind
In April 2020, Google DeepMind published a research paper on the learning method of 'AI capable of designing chip layouts' on the preprint server arXiv, and published the paper in the scientific journal Nature in June 2021. On September 26, 2024, additional information on the paper was published in Nature, and at the same time, the chip layout design AI was named ' AlphaChip '.
An IC chip is made up of multiple blocks connected by very fine wires, and the placement of each block must meet complex design requirements. For this reason, a layout design by a human would take weeks or months. On the other hand, AlphaChip can complete the layout design in just a few hours.
AlphaChip takes a similar approach to the Go AI AlphaGo and the game AI AlphaZero , treating chip layout design as a game and executing the process. Specifically, it uses a learning method that involves arranging circuit components one by one in blank squares and rewarding the user based on the quality of the resulting layout.
It is also possible to learn and generalize the relationships between components from existing IC chips and use them in other chip layout designs. Click on the images below to see videos of 'Designing the RISC-V processor 'Ariane' without learning from existing IC chips' (left) and 'Designing the RISC-V processor 'Ariane' after learning the design of 20 TPUs' (right). When you play the video, you can see that learning from existing chips allows for quicker layout determination. As AlphaChip learns more chip layouts, it becomes possible to process faster.
AlphaChip is already being used in actual chip design, and the AlphaChip is used in three models of Google's proprietary AI-specialized chip 'TPU' that were released after 2020: ' TPU v5e ', ' TPU v5p ', and ' Trillium '. In addition, the AlphaChip was also used in Google's proprietary Arm processor 'Axion' announced in April 2024. In addition, Google also provides the AlphaChip to other companies, and MediaTek has already adopted the AlphaChip in the development of its smartphone chip ' MediaTek Dimensity 5G '.
The graph below shows the number of blocks in 'TPU v5e', 'TPU v5p' and 'Trillium' that have their layout designed with AlphaChip. It can be seen that designs using AlphaChip are on the rise.
In addition, the graph below shows the reduction rate of wiring length when comparing AlphaChip design with human design. The reduction rate is gradually increasing.
The source code for AlphaChip is available at the following link:
GitHub - google-research/circuit_training
https://github.com/google-research/circuit_training
Related Posts:
