This week, Google announced the AlphaChip reinforcement learning method for designing chip layouts. AlphaChip AI promises to significantly speed up chip floorplan designs and make them more optimized for performance, power, and area. Reinforcement learning techniques, which are now publicly available, helped design Google’s Tensor Processing Units (TPUs) and have been adopted by other companies, including MediaTek.
Chip design layout, or floorplanning, has traditionally been the longest and most labor-intensive phase of chip development. In recent years, Synopsys has developed AI-assisted chip design tools that can accelerate development and optimize chip floorplans. However, these tools are quite expensive. Google hopes to democratize this AI-assisted approach to chip design to some degree.
Currently, it takes humans about 24 months to design a floorplan for a complex chip such as a GPU. Floorplanning for something less complex can take months, and the design team is usually so critical that it can end up costing millions of dollars. Google says AlphaChip accelerates this timeline, allowing you to create chip layouts in just a few hours. Moreover, its design is said to be superior as it optimizes power efficiency and performance. Google also demonstrated graphs showing wire length reductions on various versions of TPU and Trillium compared to human developers.
(Image source: Google)
AlphaChip uses a reinforcement learning model where an agent performs actions in a preconfigured environment, observes the results, and learns from these experiences to make better choices in the future. For AlphaChip, the system views chip floorplanning as a kind of game where you place circuit components one at a time on a blank grid. The system improves as it solves more layouts by using graph neural networks to understand relationships between components.
Since 2020, AlphaChip has been used to design Google’s proprietary TPU AI accelerators that power many of Google’s large-scale AI models and cloud services. These processors run Transformer-based models that power Google’s Gemini and Imagen. AlphaChip improves on the design of previous generations of TPUs, including the latest 6th generation Trillium chips, ensuring higher performance and faster development. Still, both Google and MediaTek rely on AlphaChip for a limited set of blocks, with human developers still doing the bulk of the work.
(Image source: Google)
To date, AlphaChip has been used to develop a variety of processors widely used in various smartphones, including Google’s TPU and MediaTek’s Dimensity 5G system-on-chip. As a result, AlphaChip can be generalized across different types of processors. Google says AlphaChip is pre-trained on a wide range of chip blocks and is able to generate increasingly efficient layouts as it practices more designs. Human experts learn, and many learn fast, but the pace of machine learning is orders of magnitude faster.
Expanding the use of AI in chip development
Google says AlphaChip’s success has sparked a new wave of research using AI at various stages of chip design. This includes extending AI techniques into areas such as logic synthesis, macro selection, and timing optimization, which Synopsys and Cadence already offer at great expense. Google said researchers are also studying how AlphaChip’s approach can be applied to further stages of chip development.
“AlphaChip has inspired a whole new field of research in reinforcement learning for chip design, across design flows from logic synthesis to floorplanning, timing optimization, and more,” Google’s statement reads. There is.
Looking to the future, Google sees AlphaChip as having the potential to revolutionize the entire chip design lifecycle. From architecture design to layout to manufacturing, AI-driven optimization has the potential to make chips faster, smaller (and thus cheaper), and more energy efficient. For now, Google’s servers and MediaTek Dimensity 5G-based smartphones are benefiting from AlphaChip, but the application could spread to almost everything in the future.
Future versions of AlphaChip are already in development, so stay tuned for more AI-powered chip designs.
