Libraries of Approximate Circuits: Automated Design and Application in CNN Accelerators

Mrázek, Vojtěch; Sekanina, Lukáš; Vašíček, Zdeněk

doi:10.1109/jetcas.2020.3032495

Cited by 48 publications

(32 citation statements)

References 39 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The delay of a cell is modeled as a function of its input transition time and capacitive load on the cell's output. The library was later extended to EvoApproxLib and EvoApproxLib-LITE [55].…”

Section: A Gate Level Approximationsmentioning

confidence: 99%

“…The EvoApproxLib was later extended with evolved approximate implementations of up to 128-bit adders, up to 32bit multipliers, dividers, Multiply&Accumulate (MAC) circuits, and 8 × k-bit multipliers, k ∈ {4, 5, 6, 7} [55,40,51]. Each approximate circuit is fully characterized in terms of Fig.…”

Section: A Gate Level Approximationsmentioning

confidence: 99%

“…1: The MAE-power plot for 8-bit approximate multipliers taken from the EvoApproxLib-LITE (black points), EvoApproxLib (grey points), and EvoApprox8b (red points) are compared with broken array multipliers (green points), truncated multipliers (blue points) and other selected approximate multipliers from lpAClib [57] and [56]. Adopted from [55].…”

Section: A Gate Level Approximationsmentioning

confidence: 99%

“…A lot of effort is also invested to obtaining suitable approximate implementations of multipliers because the multiplication and the MACs are dominating arithmetic operations in CNNs [62]. For CNNs with quantized weights, the 8-bit and 8xk-bit approximate multipliers are evolved by CGP with the goal to minimize WCE and power consumption of resulting multipliers [70,55]. Experiments reveal that the inexact multiplication by zero leads to the accumulation of errors in CNNs and poor classification accuracy.…”

Section: A Evolutionary Approximation Of Components Of Nnsmentioning

confidence: 99%

“…Hence, an additional constraint was formulated for CGP: if at least one of the operands is zero, the product must be precisely zero. The CNNs utilizing evolved approximate multipliers satisfying this constraint show good error-energy trade-offs for ResNet networks on CIFAR-10 dataset [55]. These evolved multipliers are also used by ALWANN tool that provides highlyoptimized implementations of CNNs for custom low-power accelerators in which the number of computing units is lower than the number of layers [71].…”

Section: A Evolutionary Approximation Of Components Of Nnsmentioning

confidence: 99%

See 4 more Smart Citations

Evolutionary Algorithms in Approximate Computing: A Survey

Sekanina

2021

JICS

Self Cite

View full text Add to dashboard Cite

In recent years, many design automation methods have been developed to routinely create approximate implementations of circuits and programs that show excellent trade-offs between the quality of output and required resources. This paper deals with evolutionary approximation as one of the popular approximation methods. The paper provides the first survey of evolutionary algorithm (EA)-based approaches applied in the context of approximate computing. The survey reveals that EAs are primarily applied as multi-objective optimizers. We propose to divide these approaches into two main classes: (i) parameter optimization in which the EA optimizes a vector of system parameters, and (ii) synthesis and optimization in which EA is responsible for determining the architecture and parameters of the resulting system. The evolutionary approximation has been applied at all levels of design abstraction and in many different applications. The neural architecture search enabling the automated hardware-aware design of approximate deep neural networks was identified as a newly emerging topic in this area.

show abstract

Section: A Gate Level Approximationsmentioning

confidence: 99%

Section: A Gate Level Approximationsmentioning

confidence: 99%