TLDR
Listen to an AI-generated podcast of the seminal 2020 paper that introduced the chip designer:
Key takeaways:
- Google’s research division, DeepMind, officially gave a name to its AI approach that can optimize the design of chips in hours rather than weeks or months for humans. It’s called AlphaChip.
- DeepMind also released a pre-trained model for others to use and customize, called the pre-trained model checkpoint.
- AlphaChip uses reinforcement learning to supercharge chip designing within hours, versus weeks or months for humans.
You’ve heard of AlphaGo, an AI program from Google DeepMind that was the first to defeat a human champion of the game Go.
Then there was AlphaFold, an AI program that solved a 50-year-old medical challenge of predicting protein structures.
Now there’s AlphaChip, which is an open-source AI approach that can optimize the design of chips in hours rather than weeks or months for humans.
While Google has been using this method in the last three generations of its enhanced AI chips or accelerators, Tensor Processing Units (TPU)s, the company has now released a pre-trained model checkpoint, sharing its model weights and also officially giving the method a name: AlphaChip.
The pre-trained model is trained on 20 TPU blocks, which is a starting point for model training and fine-turning.
Google said AlphaChip uses reinforcement learning – which rewards or punishes the machine for taking the right or wrong actions – to generate “superhuman” chip layouts that optimizes power, performance and area. These chips are deployed in areas from data centers to mobile phones and more.
A chip has many interconnected blocks, with layers of circuit components interconnected by thin wires. Because of the complexity, chip designers have tried to automate this planning process for the last 60 years, according to the company.
The DeepMind team created AlphaChip by instructing it to approach chip designing as a game. From a blank grid, AlphaChip puts circuit components in one at a time. If it’s a quality design, it gets a reward.
Importantly, the company said a novel ‘edge-based’ graph neutral network lets AlphaChip learn the relationships between interconnected chip components and to generalize across chips. This lets AlphaChip learn with each layout it designs.
Read the 2020 paper, 2021 paper in Nature and update.
