Google Researchers Create AI-ception with an AI Chip That Speeds Up AI
Reinforcement learning algorithms may be the next best thing since sliced bread for engineers looking to improve chip placement.
Researchers from Google have created a new algorithm that has learned how to optimize the placement of the components in a computer chip, so as to make it more efficient and less power-hungry.
Typically, engineers can spend up to 30 hours configuring a single floor plan of chip placement, or chip floor planning. This complicated 3D design problem requires the configuration of hundreds, or even thousands, of components across a number of layers in a constrained area. Engineers will manually design configurations to minimize the number of wires used between components as a proxy for efficiency.
Because this is time-consuming, these chips are designed to only last between two and five years. However, as machine-learning algorithms keep improving year upon year, a need for new chip architectures has also arisen.
the algorithm automatically produced hundreds of thousands of new designs, mwithin a fraction of a second, and evaluated them using the reward function. Over time, it converged on a final strategy for placing chip components in an optimal wayhttps://t.co/oTYrPx8lHj— AFENTIS FORENSICS (@afentis) March 30, 2020
Facing these challenges, Google researchers Anna Goldie and Azalia Mirhoseini, have looked into reinforcement learning. These types of algorithms use positive and negative feedback in order to learn new and complicated tasks. Thus, the algorithm is either "rewarded" or "punished" depending on how well it learns a task. Following this, it then creates tens to hundreds of thousands of new designs. Ultimately, it creates an optimal strategy on how to place these chip components.
After their tests, the researchers checked their designs with the electronic design automation software and discovered that their method's floor planning was much more effective than the ones human engineers designed. Moreover, the system was able to teach its human workers a new trick or two.