The prevalence of machine learning in recent times has had a dramatic impact on the way we design and use mobile and IoT systems. With growing concerns for privacy and quality-of-service, much of the computation that has been traditionally done in the cloud is now moving into edge devices, with far fewer resources available. This in turn increases pressure on the design and development phases, with significant improvements expected in each hardware iteration. Short design cycles require architects to rapidly explore the design space, but by definition, cycle-accurate simulators, which are the primary tool, contradict this requirement. Furthermore, GPUs operate as accelerators linked to a CPU and a tightly coupled software stack, however existing, detailed, GPU simulators are not fast or flexible enough to execute complete multi-kernel applications with heavy CPU-GPU interaction. We argue that for early design space exploration, cycle-accurate simulation is unnecessary, and can be replaced with models exhibiting strong correlation. Instead of implementing a cycle-level model, we propose a multi-phased approach to simulation, with a fast, full-system, functional simulation backed by an offline trace-based approach. By limiting the detail of both the functional and trace based models, we are able to make performance predictions that correlate strongly with real results, and are characterized by near-perfect rank correlation - both valuable metrics for early design space exploration.
|Number of pages||1|
|Publication status||Published - 12 Aug 2020|
|Event||Workshop on Modeling & Simulation of Systems and Applications 2020 - Virtual Workshop|
Duration: 12 Aug 2020 → 12 Aug 2020
|Workshop||Workshop on Modeling & Simulation of Systems and Applications 2020|
|Abbreviated title||ModSim 2020|
|Period||12/08/20 → 12/08/20|