Participant Inputs in Area 2: Circuits, Micro-architecture and Beyond

In my opinion, one of the biggest challenges in evaluating the energy efficiency of radical architectural and microarchitectural research ideas is the lack of a fully integrated flow to enable accurate energy modeling at various levels of abstraction. Traditionally, academic researchers augmented their cycle-level simulators with an ad-hoc combination of Wattch, Cacti, and Orion. Each of these are useful tools but they were developed by different research groups with different (and possibly incompatible) implementation and technology assumptions. More recently, McPAT offers a step in the right direction with a unified environment for processor, memory, and network energy modeling of traditional general-purpose multicores. Often researchers argue that absolute energy estimates are not important. Instead, we simply need to ensure that relative comparisons using high-level models (design A uses less energy than design B) track the relative comparisons of the real designs. Unfortunately, this assumption is not always validated with real implementations of the design under consideration, and in many situations we really do care about absolute energy estimates. Current microarchitectural-level models are still usually only appropriate for preliminary exploration in a very specific part of the architectural design space. For example, many research ideas require (significantly?) increasing the complexity of the control logic, yet this is usually modeled very abstractly (if at all) in current microarchitectural-level energy models. As another example, these models are mostly targeted towards very traditional architectures and (although they can be abused) are not as useful when exploring specialized accelerators, graphics processors, reconfigurable logic, emerging technologies and/or significantly different circuit implementation techniques; yet these are the exact kind of radical architectures academia should be exploring.