Approximate computing: An emerging paradigm for energy-efficient design
read more
Citations
Approximate Computing: A Survey
Mobile Augmented Reality Survey: From Where We Are to Where We Go
Approximate computing and the quest for computing efficiency
DRUM: A Dynamic Range Unbiased Multiplier for Approximate Applications
A low-power, high-performance approximate multiplier with configurable partial error recovery
References
Stochastic Computing Systems
Low-Power Digital Signal Processing Using Approximate Adders
Survey of Stochastic Computing
Stochastic neural computation. I. Computational elements
Related Papers (5)
New Metrics for the Reliability of Approximate and Probabilistic Adders
Frequently Asked Questions (15)
Q2. What is the effect of reducing the number of support vectors on the number of dot?
Reducing the number of support vectors reduces the number of dot product computations per classification while the dimensionality of the support vectors determines the number of multiply-accumulate operations per dot product.
Q3. What is the metric of reliability for an approximate circuit?
The MED is useful in measuring the implementation accuracy of a multiple-bit adder, while the NED is a nearly invariant metric, that is, independent of the size of an adder, so it is useful when characterizing the reliability of a specific design.
Q4. What is the way to reduce the critical path delay in a multiplier?
An efficient design using input pre-processing and additional error compensation is proposed for reducing the critical path delay in a multiplier [37].
Q5. What is the metric of reliability for approximate circuits?
Error rate (ER) is the fraction of incorrect outputs out of a total number of inputs in an approximate circuit [42]; it is sometimes referred to as error frequency [33].
Q6. What is the recent discussion of a synthesis of approximate circuits?
Automated synthesis of approximate circuits is recently discussed in [41] for large and complex circuits under error constraints.
Q7. What is the importance of controlling errors?
It is also crucial to control sources of errors that have the potential to be spread and amplified within the flow of the algorithm [63].
Q8. What are the performance metrics needed to evaluate the efficacy of approximate designs?
METRICS FOR APPROXIMATE COMPUTINGIn light of the advances in approximate computing, performance metrics are needed to evaluate the efficacy of approximate designs.
Q9. What is the strategy for dealing with timing errors?
When VDD is scaled down, large magnitude timing errors are very likely to happen in the addition of small numbers with opposing signs.
Q10. What is the metric of reliability used in approximate circuits?
Due to the deterministic nature of approximate circuits, the traditional metric of reliability, defined as the probability of system survival, is not appropriate for use in evaluating the quality of a design.
Q11. What are the main objectives of the paper?
In this paper, recent progress on approximate computing is reviewed, with a focus on approximate circuit design, pertinent error metrics, and algorithm-level techniques.
Q12. What led to the idea of a dithering adder?
That led to the idea of a dithering adder (Fig. 2), useful in accumulation, in which subsequent additions produce opposite-direction bounds such that the final result has a smaller overall error variance (Fig. 3) [33].
Q13. What are the common metrics used in approximate computing?
These metrics are also applicable to probabilistic adders such as those in [46, 47, 48], and provide effective alternatives to an application-specific metric such as the peak signal-to-noise ratio (PSNR).
Q14. What is the technique used to process high-frequency small-magnitude operands?
The introduced technique uses an adder with a bit-width smaller than required by other considerations to process high-frequency small-magnitude operands.
Q15. What is the effect of the number of support vectors on the quality of the algorithm?
It is found that the number of support vectors correlates well with the quality of the algorithm while also impacting the algorithm’s energy consumption.