DeepMind has revealed research on AlphaChip, a cutting-edge artificial intelligence model that operates quietly in the realm of on-chip circuits. It has been quietly utilized with the latest models of TPUs, including the Google Axion CPU and chips from other companies like MediaTek Dimensity 5G.
According to DeepMind, the current chip design process is highly complex, which is why they have introduced the concept of reinforcement learning for AI to learn independently through “playing games.” This is similar to AlphaGo and AlphaZero, but instead of games like chess or Go, the games are designed to simulate electronic circuit layouts.
Years ago, AlphaChip showcased its work on designing the prototype chip Ariane RISC-V without the need for extensive chip knowledge. Google then expanded its use to TPUs starting in 2020, including Trillium and TPU v6.
The capabilities of AlphaChip have now surpassed human capabilities. Chips designed by AlphaChip have a total wirelength that is lower than those designed by humans. This difference is at 3.2% in TPU v5e, 4.5% in TPU v5p, and 6.2% in Trillium, indicating that the more complex the circuit, the more AlphaChip’s value shines through.
TLDR: DeepMind has unveiled AlphaChip, an advanced AI model for on-chip circuitry, which has surpassed human design capabilities and is being used in the latest TPUs.
Leave a Comment