A team of researchers from Google has divulged a new AI Chip model that can come up with multifaceted chip designs in hours, a heavy complicated task that usually takes months for human engineers to finish.
Google is using AI machine learning to design its next-generation AI chips. These machine-learning algorithm designs are superior to the designs created by humans. Also, with the help of AI and machine learning, chips can be designed and generated at a faster rate as compared to those created by humans. AI chip can accomplish this work in under six hours where a human takes months, says the Tech Giant. AI, in other words, is helping accelerate the future of AI development.
For a long period of time, Google has been working on how to make use of AI machine learning to design chips. For the first time, its research has been finally applied to a commercial product. This product is an upcoming version of Google’s own TPU (Tensor Processing Unit) chip, which is augmented for AI computation.
This work has huge insinuations for the chip industry, says Goggle engineers. This initiative will allow companies to discover the possible architectural space for upcoming designs and seamlessly modify chips for definite workloads.
A dataset of 10,000 chip layouts is used by the researchers to feed a machine learning model. Only in six hours, the model could produce a design that enhances the position of the different components on the chip and also construct a final layout that gratifies operational requirements like power efficiency and execution speed.
The method is so successful that Google has already used the model to create its next-generation TPUs that run in the company’s data center to enhance the action of various AI applications.
Shortage in microchips
Modern chips comprise billions of different elements laid out and connected on a piece of silicon which is the size of a fingernail. For instance, a single processor will typically contain tens of millions of logic entries, which are called standard cells, and thousands of memory blocks, known as macroblocks, and then these have to be wired together. The position of standard cells and macroblocks on the chip is vital to determine the swiftness in the transmission of signals on the chip. This is the reason why many engineers focus on improving the chip’s layout. It starts with a process called floorplanning (placing the larger macroblocks) keeping in mind the standard cells and then placing the wired cells in the remaining place.
The number of layouts possible for a macroblock is massive. According to Google researchers, probably, there are ten to the power 2,500 different configurations possible to put to the test. Each iteration can take up to several weeks. This whole process of floorplanning is the most complex task. For several years researchers have failed to come up with technology that has the potential to remove the burden of floorplanning for engineers.
Engineers designing chips can depend on computer software to back them in such a complex task of floorplanning, but it still takes many months to look for the best way to accumulate components on the device.
As predicted by Moore Law, the number of transistors on a AI chip is getting doubled every year, therefore the challenge is now getting higher. The engineers are facing a shortage of time to meet tight schedules.
Today chips are used in everything starting from Play Stations and washing machines to toothbrushes and alarm clocks. But there is not enough supply to go around and leading to a shortage of AI chip. With time demand will be high and still supply will remain constrained.
This shortage has been caused by inadequate capabilities on the level of fabrication, as a result, there is a need to reduce the time needed to design microchips.
A Solution to the crisis
Therefore, Google’s successful initiative to automate floorplanning could be ground-breaking. Google has managed to effectively reduce the time needed to design microchips. This will be a huge help in speeding up the process of supply and omit the crisis of chip shortage.