Performance and power estimation using datamining through machine learning techniques

Abstract

The development of fast and efficient processors is an inevitable requirement in today’s automated world. New architectures that are robust and can handle complex real-time applications have to be developed. But how to find the optimum architecture and the hardware configurations in the shortest and accurate way is an open question. If the configurations must be changed after the prototype is created since it does not satisfy the customer requirements, then it will become a tedious and time-consuming process and the whole flow will need to get repeated again which is highly undesirable. Therefore, an early and exact determination of an efficient processor architecture is needed before the hardware development even starts. Modern processing systems with heterogeneous components have numerous configuration and design options such as the number and types of cores, frequency, and memory bandwidth. Hardware architects are confronted with hundreds of design parameters which can be combined in many arbitrary ways to develop new architectures. Different hardware and software configuration parameters should be evaluated by estimating the power and performance corresponding to each set without the availability of the real hardware. This highlights the importance of rapid performance and power estimation mechanisms. In this work, we propose a method to estimate power and performance of different workloads for a hardware configuration using machine learning techniques. After training the model with the data from previous generation processors, we will be able to predict the performance and power of the new generation processors with different hardware and workload features associated with that processor without having to run the simulation. This can help the architecture design team to select the best architectural configuration and build the design with optimal power and performance in a faster way to decrease turnaround times in the product lifecycle and increase the product goodness.