Tests
Our artificial intelligence method has accelerated and optimized chip design, and its superhuman chip arrays are used in hardware around the world
In 2020 we released pre-print we present our novel reinforcement learning method for chip design, which we will discuss later published in Nature AND open source.
Today we are here publishing the Nature supplement which describes more about our method and its impact on the field of chip design. We also publish pre-trained checkpointsharing the weight of the model and announcing its name: AlphaChip.
Computer chips have fueled extraordinary advances in artificial intelligence (AI), and AlphaChip is giving back by using AI to accelerate and optimize chip design. This method was used to design superhuman chip systems in the last three generations of Google’s custom AI accelerator, Tensor processing unit (TPU).
AlphaChip was one of the first reinforcement learning approaches used to solve real-world engineering problems. It generates superhuman or comparable chip layouts in hours rather than weeks or months of human effort, and its layouts are used in chips all over the world, from data centers to cell phones.
“
AlphaChip’s groundbreaking AI approach is revolutionizing a key phase of chip design.
SR Tsai, senior vice president of MediaTek
How AlphaChip works
Designing a chip layout is not a straightforward task. Computer chips are made up of many interconnected blocks, with layers of circuit elements connected by incredibly skinny wires. There are also many complicated and interrelated design constraints that must be met simultaneously. Due to the sheer complexity of integrated circuits, IC designers have tried strenuous for over sixty years to automate the IC planning process.
Like AlphaGo and AlphaZero, which learned to master Go, chess, and shogi, we built AlphaChip to approach chip floor planning as a kind of game.
Starting with an empty grid, AlphaChip places one circuit component at a time until all components are placed. It is then rewarded based on the quality of the final layout. A novel edge-based graphical neural network enables AlphaChip to learn the relationships between interconnected chip components and generalize them across different chips, allowing AlphaChip to improve any chip it designs.
Left: Animation showing AlphaChip deploying an open-source Ariane RISC-V processor with no prior experience. Right: Animation of AlphaChip placing the same block after practicing 20 TPU designs.
Using artificial intelligence to design chips that accelerate Google’s artificial intelligence
AlphaChip has generated the superhuman chip arrays used in every generation of Google TPUs since its publication in 2020. These chips enable massive scaling of AI models powered by Google technology Transformer architecture.
TPUs are at the heart of our powerful generative AI systems, starting with huge language models such as Twinsfor image and video generators, Imagen and Veo. These AI accelerators also form the heart of Google’s AI services available external users via Google Cloud.
A row of supercomputers with the Cloud TPU v5p AI accelerator in a Google data center.
To design TPU chips, AlphaChip first practices on a diverse range of chip blocks from previous generations, such as on-chip and inter-chip network blocks, memory controllersAND data transport buffers. This process is called initial training. We then run AlphaChip on the current TPU blocks to generate high-quality chips. Unlike previous approaches, AlphaChip gets better and faster as it solves more cases of the chip placement task, just as experts do it.
With each novel generation of TPU, including our latest Trillium (sixth generation) AlphaChip designed better chip layouts and provided more of the overall floor plan, speeding up the design cycle and providing higher performance chips.
Bar chart showing the number of chip blocks designed by AlphaChip across three generations of Google’s Tensor Processing Units (TPUs), including v5e, v5p, and Trillium.
Bar chart showing AlphaChip’s average wire length reduction across three generations of Google’s Tensor Processing Units (TPUs) compared to targets generated by the TPU physical design team.
AlphaChip’s broader impact
AlphaChip’s influence can be seen in its applications within Alphabet, the research community and the chip design industry. In addition to designing specialized AI accelerators such as TPU, AlphaChip has generated circuits for other chips at Alphabet such as Google Axion processorsour first general-purpose ARM-based data center processors.
Third-party organizations also deploy and utilize AlphaChip. For example, MediaTek, one of the world’s leading chip design companies, has expanded AlphaChip to accelerate the development of its most advanced chips – such as Dimenity 5G flagship used in Samsung mobile phones – while improving power, efficiency and chip area.
AlphaChip has caused an explosion of AI work in chip design and has been extended to other critical stages of chip design, such as logical synthesis AND macro selection.
“
AlphaChip has inspired a whole novel line of reinforcement learning research in chip design, cutting through the design flow from logic synthesis to floor planning, timing optimization, and beyond.
Professor Siddharth Garg, Tandon School of Engineering, Modern York University
Creating the chips of the future
We believe AlphaChip can optimize every step of the chip design cycle, from computer architecture to manufacturing, and transform chip design for custom hardware found in everyday devices such as smartphones, medical equipment, agricultural sensors and more.
Future versions of AlphaChip are currently in development, and we look forward to working with the community to continue revolutionizing this area and creating a future where chips are even faster, cheaper, and more energy effective.