Roteamento global de circuitos vlsi utilizando aprendizado por reforço

Abstract

Nowadays, there is an increase demand for compact electronic devices such as mobile phones, smartwatches, and so on. There is also an inverse relationship between device’s size and complexity of its integrated circuit (IC). To deal with issues related to the increase in the circuit complexity, the design process starts with layout definition called routing. The routing stage is divided into global and detailed. The global routing is the most complex stage of the IC physical design, since indeed it is a NP-Complete problem. Classic solutions are based on meta-heuristic algorithms. This paper presents a new approach to global routing, namely Router-DQN, which refines the A* routing by using deep reinforcement learning techniques. We evaluated the presented router algorithm with two main metrics: congestion (routing demand exceeding the routing capacity) and the average length of the networks. The results obtained indicate that the Router-DQN overcomes classical approaches in the literature.