How Google used Machine Learning to dramatically improve chip design?
The recent advancements in computer systems and hardware have revolutionized the modern computing system. With Moore’s law(which states that the number of transistors on a microchip doubles every year) diminishing due to new ways of measuring processing power, the need for specialized hardware is growing exponentially to match the modern-day computing demands.
However, the microchips are in shortage to match the ongoing computing needs of people. With the Covid-19 pandemic, the requirement for electronic goods has almost doubled in every house. The increased use of laptops, PCs, smartphones has caused a shortage of microchips.
The microchip manufacturers worldwide could not match the growing computing needs of the people. Nevertheless, microchip design is a tedious process that often demands intense effort from engineers.
Microchips are made of 100 layers of interconnected patterns on a silicon wafer, making microchip manufacturing a challenging process. Even with highly skilled engineers, it takes a few months for them to design microchips.
But now, the tech giant Google has transformed the chip designing process by introducing Machine Learning(ML). With ML, Google has shortened the chip design process drastically while improving the design dramatically. Let’s see how Google made it possible.
What Google Scientists Say About Using Machine Learning (ML) to design Microchips?
Google Scientists at Google Brain have declared introducing deep reinforcement learning techniques for floorplanning. In reality, floorplanning is a time-consuming process of arranging the placement of various components on a microchip.
Google decided to revolutionize the challenging chip design process with AI. Thus, the scientists used ML to design optimal chips for AI processors, the next generation of Tensor Processing Units, shortly.
The Google scientists clarify that the use of software in chip design is nothing new. But, the new reinforcement learning model generates superior chip floorplans automatically. These AI-designed chips are highly optimal than the human-produced chips in all prime aspects like power consumption, performance, and chip area. Plus, AI can design chips in a few hours, unlike humans do in a few months.
Thus the AI-designed chips have become the talk of the town in the computing field. Google has used ML successfully to improve the chip design dramatically. AI software can design computer chips much faster than humans can, in less than six hours.
Why Does Google Prefer ML for Chip Design?
1) It allows you to overcome the chip floorplanning problem:
Google explains that the company overcame the chip floorplanning problem by leveraging ML. Though the chip fabrication process is automated, the design is still a manual process.
Even with computer-aided design software, it still takes weeks or months to work out how to fit all the components into the available space. Google scientists confirm that by leveraging ML, we can reduce the design process in a few hours.
Ideally, the area of a microchip is a few hundred millimeters square. Designers need to accommodate various components like memory, logic, and processing units in that area. Moreover, they need to place kilometers of ultra-thin wire to connect those components.
The most demanding factor of the design process is chip floorplanning. It is all about working out where to place these components in the best possible way. Thus, chip floorplanning is often a daunting task for designers.
Google scientists utilized around 10,000 chip floorplans to train their software. The software then worked out how to create floorplans that needed minimal space, wire, and power than an engineer’s design. These AI-designed chips are ideal to use in smartphones as they come with miniaturization and consume less electricity.
2) It saves a lot of time
The AI-generated chips took less than six hours to design, unlike the regular chip that demanded months of hard work from engineers. Plus, the AI chip floorplans are highly superior to those human-designed ones in all aspects like power consumption, performance, and chip area.
Google achieved time reduction by posing chip floorplanning as a reinforcement learning problem and developed an edge-based graph convolutional neural network architecture capable of learning rich and transferable representations of the chip. Thus, the chip design process took less than six hours rather than weeks or months.
How Google used ML to Improve Chip Design Dramatically?
Any organization can derive value from AI by following Google’s approach. Machines often outsmart people in specific areas where raw computing power is more crucial than creative insights. As we all know, humans do better in creative insights. That is why various authorization processes use creative things like pattern matching to verify humans or robots, as AI can’t match patterns.
Here, Google engineers came up with an intelligent algorithm that did not send it to the design chip. Alternatively, they pre-trained agents with a set of 10,000 chip floor plans using reinforcement learning.
The agent here learns to predict the expected success rates by analyzing the past success rates. The trained agent often examines the state of the chip under development during all steps of the floor plan. It even assesses the partial floor plan built so far and applies the learned method to determine the best possible area to place the next macroblock.
Thus, ultimately, Google applied the floor plan solution in the chip design of its next-generation artificial intelligence processors. So, it is not a mere scientific experiment. Instead, it is an AI-based approach to chip design that is already benefiting the chip design process.
So, even other chip manufacturers can also do their AI research to improve their chip design processes, like Google. All you need is to have specific goals and train the agent with well-defined and structured training data. Also, as many think, AI will only augment human work and will never replace human scientists.
Final Thoughts
Google’s success with AI-enabled chips is not only about chip design. The tech giant acts as the pioneer here, giving hope to numerous other organizations from various industries to implement AI for any tedious processes.
Many industries like autonomous vehicles, 5G communications, and more will need more efficient microchips. Since the assurance is from Google itself, many other giants have already started implementing AI in their manufacturing or design processes. So, whatever industry your business belongs to, you can explore the possibility of AI in your field.
See how Google transformed the tedious microchip design process using AI technology?! Likewise, you can identify the challenging process in your niche and revolutionize it with AI.
Often people fear automation as they think it will replace their jobs. But on the contrary, it will only augment human work. In the electronics industry, companies are already planning to create the next generation of microchips.
Thus, business owners must stay updated about the current technologies, trends, and benefits. So, you can train your team with relevant skills and expertise or upskill yourself.
That will allow your organization to adapt to any new technology in the future spontaneously, enabling you to scale up your business and boost profits.
