Our AI method accelerated and optimized the structure of the systems, and its superhuman systems are used in equipment around the world
In 2020 we published Reprint Introduction of our innovative method of learning reinforcement to designing systems of systems, which later Published in nature AND Open acquisition.
Today we are Publishing the additive This describes more about our method and its impact on the field of system design. We also release pre -trained checkpointSharing weight and announcing its name: Alphachip.
Computer chips have fueled unusual progress in artificial intelligence (AI), and Alphachip returns favor by using artificial intelligence to accelerate and optimize systems design. The method has been used to design superhuman systems in the last three non -standard generations of AI Acelerator Google, Tensor processing unit (TPU).
Alphachip was one of the first approaches to learning the strengthening used to solve the engineering problem in the real world. It generates superhuman or comparable chip systems in hours, not for weeks or months of human effort, and its systems are used in systems around the world, from data centers to mobile phones.
“
The AI Alphachip breakthrough is revolving the key phase of layout design.
SR Tsai, senior vice president of MediaTek
How Alphachip works
Designing the system system is not a simple task. Computer chips consist of many connected blocks, with layers of circuit components, all connected with extremely thin wires. There are also many complex and intertwined design restrictions, which all should be met at the same time. Due to their complexity, systems designers have tried to automate the floor planning process for over sixty years.
Like Alphago and Alphazero, who learned to control him, Chess and Shogi games, we built Alphachip to get closer to planning floors as a game.
Starting from the empty mesh, Alphachip places one element of the circuit at once until it ends to place all the elements. Then it is awarded based on the quality of the final system. The innovative “edge -based” neural network allows Alphachip to learn the relationship between combined systems components and generalize between systems, allowing Alfachip to improve with every project he designs.
On the left: Animation showing Alphachipa placing Open Source, Ariane RISC-V CPU, without prior experience. On the right: Animation showing that Alphachip places the same block after exercising in 20 TPU projects.
Using artificial intelligence to design chips acelerator Google AI
Alphachip has generated superhuman systems used in every generation of TPU Google since its publication in 2020. These systems allow a massive increase in AI models based on Google Transformer architecture.
TPU lie in the heart of our powerful AI generative systems, from large language models, like TwinsFor image and video generators, Imagen and VEO. These AI accelerators also lie in the heart of AI Google services and are available To external users via Google Cloud.
The Cloud TPU V5P government AI AI supercomputers in the Google Data Center.
To design TPU systems, Alphachip, first practices on a variety of system blocks from previous generations, such as Network blocks on the chip and between the chipIN Memory controllersAND Data transport buffers. This process is called initial training. Then we launch Alphachip on current TPU blocks to generate high quality systems. Unlike earlier approaches, Alphachip becomes better and faster because it solves more cases of the task of placing systems, just like human experts.
With each new generation of TPU, including our latest Trillium (6. Generation), Alphachip designed better systems and provided more general floor plan, accelerating the design cycle and giving higher performance systems.
The bar chart showing the number of chip blocks designed by Alphachip in three generations of TENSOR processing devices (TPU), including V5E, V5P and Trillium.
A bar chart showing the average reduction of Alphachip lock length in three generations of Google (TPU) processor processing units, compared to places generated by the TPU physical design team.
Wider influence of Alphachip
The influence of Alphachipa can be seen using its applications in Alphabet, the research community and the chip design industry. In addition to the design of specialist AI accelerators, such as TPU, Alphachip generated systems for other alphabet systems such as Axion Google processorsOur first processors from the general medium -based data.
External organizations also accept and build on Alphachip. For example, MediaTek, one of the best design companies in the world, has expanded Alphachip to speed up the development of their most advanced chip systems, while improving the power, performance and area of chip.
Alphachip caused an explosion of work on artificial intelligence to design systems and has been expanded to other critical stages of chip design, such as Logical synthesis AND Macro selection.
“
Alphachip has inspired a completely new line of research on learning to learn in chip design, crossing the flow of the project from logical synthesis to floor planning, time optimization and more.
Professor Siddharth Garg, Nyu Tandon School of Engineering
Creating the future tokens
We believe that Alphachip has the potential to optimize each stage of the design cycle, from computer architecture to production – and transform chip design for non -standard equipment found in daily devices such as smartphones, medical equipment, agricultural sensors and many others.
Future versions of Alphachip are now in development and we are looking forward to cooperation with the community to continue revolutionizing this area and lead to the future in which the tokens are even faster, cheaper and more efficient.
Thanks
We are very grateful to our amazing co-authors: Mustafa Yazgan, Joe Wenjie Jiang, Ebrahim Songhori, Shen Wang, Young-Joon Lee, Eric Johnson, Omkar Pathak, Azade Nazi, Jiwoo Pak, Andy Tong, Kavya Srinivasa, William Hang, and Jeff Dean.
We especially appreciate Joe Wenjie Jiang, Ebrahim Songhori, Young-Joon Lee, Roger Carpenter and Sergio Guadarram's constant efforts to determine this impact on production, Quoc V. Le for research advice and mentoring and our older author Jeff Dean for his support and discussions about deep technical.
We also want to thank Ed Chi, Zoubin Ghahramani, Koray Kavukcuoglu, Dave Patterson and Chris Manning for all their advice and support.