Disciplined Approximate Computing for Energy Efficiency and Resilience
Abstract
The work in the proposed study has four core components: identifying approximation opportunities in the applications in question; mapping out the space of potential approximation techniques; compiling models of accuracy and efficiency trade-offs for each technique; and combining the application model with the approximation technique models and produce quality-versus-efficiency gain plots. Below we list each task and subtask. In Step 1, we will collect applications and test cases, assess their error resilience, profile them looking for hot code, computer patterns and memory usage. Then in Step 2 we will compile a set of techniques for approximate computing, including deterministic and nondeterministic functional units and approximate accelerators, approximate storage, and approximate communication techniques. Then in Step 3 we will match application opportunities found in Step 1 and match them with Step 2. We will then, in Step 4 derive and build analytical and/or simulation models for the techniques. Finally, in Step 5, we will organize a workshop for the key players in approximate computing write a report on the study.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 02, 2023
- Accession Number
- AD1221981
Entities
People
- Luis Ceze
- Mark H. Oskin
Organizations
- University of Washington