| dc.description.abstract | In modern System On Chip (SOC) designs, the Power Delivery Network (PDN) plays a crucial role. Its primary function is to efficiently distribute power to every component on the chip, ensuring that each component receives an ample and stable power supply. IR-drop refers to the voltage drop phenomenon from the power source to the component due to the inherent resistance characteristics of the physical structure of the Power Delivery Network (PDN). Insufficiently designed PDN can lead to excessive IR-Drop, and such excessive IR-drop is often accompanied by reliability issues in the circuit, including timing violation and functional failure. This makes the IR-drop phenomenon an issue in IC design that cannot be ignored.
The planning of the PDN typically occurs in the floorplan stage before placement. Therefore, without placement, CTS, and routing information, it is challenging to design a perfect PDN at the early stage of the physical design flow. Hence, IR-drop is a common issue frequently encountered throughout the entire physical design flow. In many previous studies on improving IR-drop, a common approach is to adjust the physical structure of the PDN itself. This can be achieved by globally or locally increasing the width or density of power stripes. This technique effectively reduces the equivalent resistance from the power source to the component, thereby reducing the IR-Drop between these two points
However, as the P&R (Placement and Routing) stages progress, the available routing space decreases. Adjusting the physical structure of PDN within limited space may potentially impact the existing P&R results. In other words, adopting the aforementioned method would require multiple iterations of P&R. For industrial-scale designs with millions or even tens of millions of standard cells, despite the assistance of state-of-the-art EDA tools, both placement and routing still demand dozens of hours or even days. In the scenario where IR-drop is diagnosed in the late stages of P&R, going back to modify the physical structure of PDN would incur significant time costs, implying that the approach of adjusting the physical structure of the PDN itself to mitigate IR-Drop will inevitably encounter bottlenecks.
Therefore, this work proposes an approach to add extra power wires at the post-routing stage to improve IR-drop, with the help of reinforcement-learning. Compared to most previous works, our approach offers the following advantages: (1) By efficiently utilizing remaining routing spaces to add power wires, it avoids the need to adjust the main structure of the PDN, thus maximizing the preservation of existing P&R results. (2) In contrast to heuristic algorithms, reinforcement learning continuously updates strategies through trial and error, providing a better chance of finding the global optimal solution. | en_US |