Unlike prior approaches, AlphaChip is a learning-based method, meaning that it becomes better and faster as it solves more instances of the chip placement problem. This pre-training significantly improves its speed, reliability, and placement quality, as discussed in the original Nature article and in a follow-up study at ISPD 202218. Incidentally, pre-training also gives rise to the impressive capabilities of large language models like Gemini19 and ChatGPT20 (the “P” stands for “pre-trained”).
Since publication, we have open-sourced a software repository21 to fully reproduce the methods described in our paper. External researchers can use this repository to pre-train on a variety of chip blocks and then apply the pre-trained model to new blocks, as was done in our original paper. As part of this addendum, we are also releasing a model checkpoint pre-trained on 20 TPU blocks22. For best results, however, we continue to recommend that developers pre-train on their own in-distribution blocks18, and provide a tutorial on how to perform pre-training with our open-source repository23.
Obviously, not performing any pre-training at all circumvents the learning aspects of our method by removing its ability to learn from prior experience.