ImpBench revisited: An extended characterization of implant-processor benchmarks

Christos Strydis,  Dhara Dave,  Georgi Gaydadjiev
Delft University of Technology


Abstract

Implants are nowadays transforming rapidly from rigid, custom-based devices with very narrow applications to highly constrained albeit multifunctional embedded systems. These systems contain cores able to execute software programs so as to allow for increased application versatility. In response to this trend, a new collection of benchmark programs for guiding the design of implant processors, ImpBench, has already been proposed and characterized. The current paper expands on this characterization study by employing a genetic-algorithm-based, design-space exploration framework. Through this framework, ImpBench components are evaluated in terms of their implications on implant-processor design. The benchmark suite is also expanded by introducing one new benchmark and two new stressmarks based on existing ImpBench benchmarks. The stressmarks are proposed for achieving further speedups in simulation times without polluting the processor-exploration process. Profiling results reveal that processor configurations generated by the stressmarks follow with good fidelity - except for some marked exceptions - ones generated by the aggregated ImpBench suite. Careful use of the stressmarks can seriously reduce simulation times up to x30, which is an impressive speedup and a good tradeoff between DSE speed and accuracy.