Computing
1
2
3
4
5
Download this chapter (PDF)
Trade-offs in accuracy with approximate computing

There are computer applications that do not require high (computational) accuracy and precision, or work with information that already contains uncertainty. Think of applications like neural networks, signal processing, and localisation and mapping for which an approximate result is sufficient to meet both the requirements and design goals in the computer architecture.

‍

Design goals can be traded, such as performance, power, and energy efficiency. Techniques available for approximate computing include quantisation, rounding, truncation, and reduction of the number of bits.

‍

Approximate computing is also achieved at the software level with dedicated algorithms, and at the hardware level by using approximate hardware, or via a combination of both. Therefore, approximate computing is gaining more traction, leading to trade-offs in accuracy in exchange for other improvements in computational objectives.

Maturity

WATCH
PLAN
ACT
Show the other ACT trends
More diversified access to large AI models Increase in co-evolution between hardware and AI Collaboration between humans and AI Democratisation of XR content creation using generative AI Public investments in XR continue to increase Standardisation of data space architectures for secure and trusted data sharing Growing significance of augmented data management Trusted digital recognitions are emerging Digital Trust Frameworks: from hierarchical to distributed trust Emerging dualuse of AI in cybersecurity More accessible privacy-enhancing technologies to share data Rising pressure to prepare for post-quantum cryptography Emergence of measures for secure Internet of Things The growing need for security protocols to work in the cloud Growing investments in the European cloud infrastructure & sovereignty More AI-driven hardware in data centres to manage workloads in the cloud Increasing need for heterogeneous and hybrid computing platforms Trade-offs in accuracy with approximate computing Fixed wireless access Fibre optic networks: strategic infrastructure for the knowledge economy Enhancing quantum computing robustness and potential Quantum sensing integration to other domains

Drivers

Impact

education

Education

  • Computer science is mostly based on determinism and exact computation. Students need to be trained to deal with approximated and not ‘exactly correct’ results. Concepts like error management and computation accuracy within known boundaries are likely to emerge both in theory and in practical sessions.
Research

Research

  • The approximate computing paradigm is reasonably mature. Optimisation based on approximate computing might be beneficial for specific applications such as neural networks and could boost the performance of several research application domains, such as in the medical sciences.
Operations

Operations

  • Generally, the use of approximate software routines does not require an update of IT infrastructure. Larger benefits could come from the use of approximated hardware, but the changes of the infrastructure required must be contrasted with the resulting benefit. Such evaluation should be done individually on a case-by-case basis.
More info about Computing Technologies?
Visit surf.nl
Link SURF icoon