Automated Implant-Processor Design

More Info


As we move towards an aging population, it is likely that an increasing number of people will require an increasing diversity of implants, but at a lower cost to the society. Also, as computer technology progresses, smaller, more powerful, and less battery intensive implants can be designed. However, present implant design methodology is highly inefficient at meeting these goals as it suffers from non-reuse of existing knowledge by relying heavily on custom designs and ASICs. The SiMS project was started with the goal of creating pre-designed, pre-tested, and pre-certified toolbox of components for biomedical implants that can be assembled in a modular fashion for various application scenarios. One of the most important components in such a tool-box is the processor. Designing such a processor is a non-trivial task and previous work has concentrated on studying the effect of changing the processor input-parameters (such as caches), one parameter at a time. The present work represents a shift in this methodology, as we now allow co-variation in all possible input parameters in order to find optimal configurations in terms of the output objectives - power, performance, and area. Towards this end, we implement ImpEDE -- "Implantable-processor Evolutionary Design-space Explorer" -- a framework that performs multi-objective optimization of processor parameters, and hence gives as output a Pareto optimal set of processors. The framework consists of a cache simulator and a cycle-accurate processor simulator running benchmarks and workloads designed for medical implants, in order to simulate the optimization objectives. A popular, highly configurable, multi-objective genetic algorithm, NSGA-II, performs the actual optimization. Supporting scripts add modularity by acting as the interface between the genetic algorithm and the simulators, enabling easy replacement with new simulators. The whole framework is parallelized such that extra computation cycles of the idle laboratory CPUs can be utilized, thereby giving a considerable speedup without requiring any special hardware. We perform experiments on the non-dominated solution fronts evolved by the framework on a sub-set of benchmarks, in order to optimize parameters of the genetic algorithm, with an aim towards speeding up convergence. We also examine the effects of changing the workload size run by the benchmarks. A solution Pareto optimal front consisting of optimal processor configurations across all benchmarks is found. This front is used as a reference in order to characterize the benchmarks in the ImpBench suite. Finally, the objective space of the reference front is compared to existing implant designs, and a set of "generic processors" are chosen such that all the existing implant applications studied can be covered.