Silicon Valley giant, Google has reportedly claimed that it has created an AI that can design chips faster and as good, if not better than humans.
According to Googlers, Anna Goldie, Azalia Mirhoseini and their colleagues, it is a deep reinforcement learning system which can develop floorplans within six hours. While human engineers and automated tools can take months to generate an optimal layout.
For the record, the AI designs the floorplan of the chip, which is the arrangement of its subsystems like CPU and GPU cores, RAM controllers, cache memory, etc. on its silicon die. Placement of the minute electronic circuits that constitute these modules can affect the power consumption and processing speed of the microchip. The signal routing and wiring required to connect all the modules is of utmost importance.
Sources have reported that Google has used the AI to produce a floorplan of next-generation TPU (Tensor Processing Unit). The company utilizes the TPU for speeding up neural networks in its search engine, public cloud, AlphaZero, AlphaGo, and other products and projects.
Reportedly, the company is effectively using machine-learning technology for optimizing the future chips that expedite machine-learning software.
The Googlers stated that their method was used for designing the next generation artificial-intelligence accelerators for Google which has the potential of saving thousands of hours of human effort for every new generation. They believe that the highly powerful AI-based hardware will stimulate advancements in AI that are expected to create a synergetic relationship between the two fields, they added.
For speeding up the floor planning stage, the company’s AI scientists reportedly developed a convolutional neural network system capable of performing all the macro block placement by itself within hours to produce an fine layout. It involves automatic placement of the standard cells in the gaps using other software.
This machine-learning based system is expected to produce an excellent floorplan rapidly and better than the above-mentioned method involving iterating and tweaking of floorplan using the existing conventional automated tools and human control.