Writing · 2024
Approximate Computing: Trading Accuracy for Efficiency
Why deliberately computing the wrong answer is a legitimate design strategy once the thermal dissipation barrier, not transistor count, is the binding constraint.
Sole-authored, Cal Poly Pomona · 2024
Approximate computing raises efficiency by giving up accuracy on purpose. The justification is a physical one: we have kept shrinking the chips, but the thermal dissipation barrier has not moved, so power — not density — is what limits what a processor can do.
Reducing the precision of operations reduces the power those operations consume, which converts directly into efficiency for any workload that can tolerate the error.
The survey covers hardware techniques, software-level approximation, and the research applications where the trade is most defensible — which is the real question, since the strategy is only sound where the application has genuine error tolerance to spend.