Special Session on Approximate Computing

Approximate Computing Circuits and Systems

Organizers:

  • Pasi Liljeberg, University of Turku, Finland
  • Anil Kanduri, University of Turku, Finland
  • Jari Nurmi, Tampere University, Finland

Rationale and scope:

Performance, implementation complexity and energy consumption have been the key optimization parameters for digital circuits and systems already for decades. Regarding energy, it is made up from the power consumption and computation time (E = P x t), where the time represents performance, and both the power and time are also dependent on the implementation complexity. There is an inherent trade-off between these optimization criteria, which is hard to break to bring the energy consumption down considerably. However, by introducing a new dimension, accuracy, to the design optimization, significant reductions in the computation complexity, power and time and thus lower energy can be achieved. The energy efficiency can even be improved by a factor of 10x-50x in the most prominent cases, which is far beyond what can be reached when sticking to the conventional and conservative best available accuracy.

There are several application areas where reducing the accuracy either end-to-end or locally is tolerable, e.g., machine learning, data mining, sensor signal processing, and audiovisual human-perceived signals. The accuracy reduction can be permanent or adjustable at run-time. Resource constrained devices such IoT and edge processors have benefitted from the energy efficiency by exploiting the accuracy trade-offs. Opportunities for approximation span vertically across the computing stack ranging from algorithms down to low level hardware. Cross layer interaction and awareness is one of the key areas for disciplined tuning of approximation to maximize the performance and energy gains while ensuring accuracy guarantees.

For this special session, we are looking for Approximate Computing solutions in the following areas (but not strictly limited to):

  • Approximate memory techniques
  • Approximating logic implementations
  • Arithmetics and number systems supporting multiple precisions
  • Approximate algorithms and software
  • Error resilience techniques allowing for approximation
  • Analog approximations of digital functionalities
  • Adaptive approximation approaches
  • Systems and applications exploiting approximate computing
  • Application/domain specific optimizations

Submission:

See the submission page to see how to submit a paper to this session. Special Session papers will undergo similar reviews as any other papers submitted to the conference.