Application area: Low Power Design
Researcher in charge: Sayandip De
Supervisor: prof. dr. Henk Corporaal, ir. Jos Huisken
Host: Technische Universität Eindhoven, The Netherlands
Secondments: Intel, Eindhoven and Inst. Of Computer Technology, Vienna Institute of Technology, Austria.
Over the past few years, the difficulty to follow Moore’s law due to the breaking of Dennard’s Scaling and the increasing influence of “Dark Silicon” has called for a change in the current computing paradigm. Researchers have been focussing on creating new devices using new materials like carbon-nanotubes, graphene etc., building newer architecture using near-memory computing and going to altogether new computing paradigms like quantum computing, neuromorphic computing etc. Even though these are interesting directions, they seem to be quite far-fetched. A key observation is that current computing systems always compute an accurate result even though all applications do not need the same degree of accuracy all the time. Incorporating this new-dimension of accuracy into current computing systems can help us achieve higher performance & energy efficiency. This idea is promising and much more realistic, which motivates us to move towards “Approximate Computing”.
The main objective of this project is to focus on the development of an approximate system-to-silicon framework which, given an application and its quality constraints, can come up with an error prone implementation on a target platform while adhering to the quality needs and also optimizing for energy/performance. The implementation should consider approximations across the different layers of design abstractions while finding the optimal trade-off between the impact of approximation on one layer upon the other. The flow for such a framework should be as follows:
- Indepth understanding and analysis of inherent application resilience across a broad range of applications. Analyzing the impact of resilience variability of factors such as input distributions, target implementation platform etc.
- Tools to automatically identify & annotate exact resilient regions in the application followed by assigning proper accuracy bounds.
- Methods to quickly assess the potential of various approximate computing techniques for a given application and given platform for implementation.
- Implementation of cross-layer approximate computing techniques followed by analysis of the impact of one technique on other and finally, optimizations at each level to get the most optimal energy efficiency/performance improvements while adhering to the application quality requirements.
- Proper synchronization between logic and data approximation techniques.
This research topic of this ESR is tightly connected to the research lines RL1 and RL2. We aim for developing a resource constraint platform with a strict energy and power budget (RL2) following a synergistic cross-layer optimization approach across the system stack.