MLNoC: A Machine Learning Based Approach to NoC Design

Modern System on Chips (SoCs) are becoming increasingly complex with a growing number of CPUs, caches, accelerators, memory and I/O subsystems. For such designs, a packet based distributed networks-on-chip (NoCs) interconnect can provide scalability, performance and efficiency. However, the design of such a NoC involves optimizing a large number of variables such as topology, routing choices, arbitration and quality of service (QoS) policies, buffer sizes, and deadlock avoidance policies. Widely varying die sizes, power, floorplan and performance constraints across a variety of different market segments, ranging from high-end servers to low-end IoT devices, impose additional design challenges. In this paper we demonstrate that there is a strong correlation between SoC characteristics and good NoC design practices. However this correlation is highly non-linear and multidimensional, with dimensions indicative of the features of the SoC, design goals and properties of the NoC. This results in a high-dimensional NoC design space and complex search process which is inefficient to solve with classic algorithms. Using a variety of real SoCs and training data sets, we demonstrate that a machine learning (ML) based approach yields near-optimal NoC designs quickly. We determine a number of SoC and NoC features, describe reduction methods, and also show that a multi-model approach yields better designs. We demonstrate that for a wide variety of SoCs, ML based NoC designs are far superior to those designed and optimized manually over years on almost all quality metrics.