OPPORTUNISTIC-DSP

Opportunistic Approximations to Break the Traditional Efficiency Limits of Flexible DSP Implementations

 Partecipanti

Mappa

 Word cloud

Esplora la "nuvola delle parole (Word Cloud) per avere un'idea di massima del progetto.

 Obiettivo del progetto (Objective)

'Worst-case design is one of the keys to practical engineering: create solutions that can always withstand all the most adverse conditions that could occur. Yet, the limitations of battery capacity, the slowing progress in CMOS technology energy efficiency, and the ever growing need for more computational power combine into a grim outlook for worst-case design in the future. This becomes even worse, when one considers the exploding non-recurring engineering costs and turns toward programmable systems—an economic necessity in most cases and yet a fundamentally less energy-efficient implementation option. This project pretends to revolutionize the way systems are design today introducing the concept of opportunistic approximations to fundamentally depart from the notion of a single worst-case design point. Opportunistic approximations enable a continuous adaptation of the processing characteristics (operation type, number and precision) to the actual operating conditions. By relaxing the processing load when possible, a significantly increase in the processing efficiency of advanced DSP (Digital Signal Processing) flexible implementations is expected. Particularly, for DSP systems running under highly varying situations. Making opportunistic approximations practical requires the combination of deep knowledge of DSP applications and low power flexible DSP architectures, expertise developed by the applicant during his PhD, with the latest design automation technology, area where the proposed host institution excels worldwide. Thus, enabling opportunistic approximations will result in major reductions of average energy consumption and average processing time while still meeting functional specifications at the input-output interface (e.g., maximal bit-error-rate for a wireless system).'

Introduzione (Teaser)

An innovative new way of building digital computers which dynamically adjusts the precision of numerical computations to reduce energy consumption without altering user experience. Even better, its applications are virtually limitless.

Descrizione progetto (Article)

Digital signal processing (DSP) refers to a host of techniques used to improve digital signals for enhanced accuracy and reliability. The world is inherently analogue, but computers (actually, analogue-to-digital converters) digitise them to use them. DSP helps to distinguish between the signal and the noise.

DSP applications are virtually limitless. They include digital wireless communication modems and smartphones, health monitoring and video and audio processing. Engineering systems are often designed to meet the worst case scenario and DSP is no exception. Processing precision is held at a very high level across all cases at the expense of energy efficiency.

A novel practical framework for dynamically reducing precision when conditions are appropriate could have substantial impact on energy use and on waste. The EU-supported project OPPORTUNISTIC-DSP developed a scheme for applications requiring very high energy efficiency where maximum benefit might be achieved by lowering numerical precision when possible.

Scientists identified a number of context-specific relaxed specifications, adapted computations for the specific context (operator, number and precision), and monitored execution conditions continuously to find opportunities for simplification. Changing the numerical precision at appropriate times has major impact in a highly dynamic computational scenario where the difference between worst case and typical case numerical precision can be enormous. When implemented on an advanced multiple input, multiple output receiver (the Long Term Evolution (LTE) developed by the Third Generation Partnership Project (3GPP)), the new approach achieved a 40 % reduction in energy consumption compared to the fixed precision implementation.

While the concept is simple, its implementation is extremely tedious since it entails multiple finite precision refinements, which are very time-consuming optimization processes. Recognising this, the team developed three approaches to enable greater automation of the design process to minimise design complexity.

Digital representations of the world have become the norm with computations routinely carried out quickly by DSP equipment and computers. OPPORTUNISTIC-DSP has made important progress toward algorithms and techniques that will facilitate smart reductions in precision through opportunistic approximations. These techniques were shown to significantly reduce energy consumption without degrading functionality and while being complementary to many other traditional energy-saving optimisation techniques. Thus, opportunistic DSP may be an important way to decrease computational load and simultaneously decrease electricity consumption and associated emissions.