SmartApprox: Learning-based configuration of approximate memories for energy-efficient execution

Filho, Jo�ão Fabrício; Felzmann, Isaías; Wanner, Lucas

doi:10.1016/j.suscom.2022.100701

Cited by 3 publications

(3 citation statements)

References 4 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Previous studies that exploit the reduced-voltage DRAM concept mainly aim at improving the energy efficiency of mobile systems (Haj-Yahya et al, 2020 ), personal computing systems (Nabavi Larimi et al, 2021 ; Fabrício Filho et al, 2022 ), and server systems (David et al, 2011 ; Deng et al, 2011 , 2012a , b ; Nabavi Larimi et al, 2021 ). This concept is also employed for minimizing the energy consumption of deep neural networks (DNNs) (Koppula et al, 2019 ).…”

Section: Resultsmentioning

confidence: 99%

EnforceSNN: Enabling resilient and energy-efficient spiking neural network inference considering approximate DRAMs for embedded systems

2022

View full text Add to dashboard Cite

Spiking Neural Networks (SNNs) have shown capabilities of achieving high accuracy under unsupervised settings and low operational power/energy due to their bio-plausible computations. Previous studies identified that DRAM-based off-chip memory accesses dominate the energy consumption of SNN processing. However, state-of-the-art works do not optimize the DRAM energy-per-access, thereby hindering the SNN-based systems from achieving further energy efficiency gains. To substantially reduce the DRAM energy-per-access, an effective solution is to decrease the DRAM supply voltage, but it may lead to errors in DRAM cells (i.e., so-called approximate DRAM). Toward this, we propose EnforceSNN, a novel design framework that provides a solution for resilient and energy-efficient SNN inference using reduced-voltage DRAM for embedded systems. The key mechanisms of our EnforceSNN are: (1) employing quantized weights to reduce the DRAM access energy; (2) devising an efficient DRAM mapping policy to minimize the DRAM energy-per-access; (3) analyzing the SNN error tolerance to understand its accuracy profile considering different bit error rate (BER) values; (4) leveraging the information for developing an efficient fault-aware training (FAT) that considers different BER values and bit error locations in DRAM to improve the SNN error tolerance; and (5) developing an algorithm to select the SNN model that offers good trade-offs among accuracy, memory, and energy consumption. The experimental results show that our EnforceSNN maintains the accuracy (i.e., no accuracy loss for BER ≤ 10−3) as compared to the baseline SNN with accurate DRAM while achieving up to 84.9% of DRAM energy saving and up to 4.1x speed-up of DRAM data throughput across different network sizes.

show abstract

Section: Resultsmentioning

confidence: 99%

EnforceSNN: Enabling resilient and energy-efficient spiking neural network inference considering approximate DRAMs for embedded systems

2022

View full text Add to dashboard Cite

show abstract

“…These errors are nondeterministic and may occur at any point on the exposed data according to a probability related to the configuration parameters, which provides control of the approximation [130]. Figure 1.2 shows the relative energy consumption collected from the execution of diverse applications with an instance of an approximate DRAM face to the error probability per access in this memory, based on an error characterization of the DRAM voltage [12] executing on a controlled environment with a median instance of error rate [30]. The energy consumption decreases as the vdd is adjusted below the guard-band margin, however, the error probability grows exponentially at each step of the supply voltage.…”

Section: Approximate Memoriesmentioning

confidence: 99%

“…Thus, no quality specifications or metrics are required for the input application, and its approximation level is determined based on the error tolerance of the training applications and the variables that influence the error rate. Our main contributions, built upon previous work [30], are:…”

Section: Chapter 4 Learning-based Configuration Of Approximate Memoriesmentioning

confidence: 99%

Software and hardware interfaces for approximate memories

Fabrício Filho

View full text Add to dashboard Cite

I would like to thank the many people that give me support during this Ph.D.To my wife, Aline Maria Gonçalves Fabrício, for walking alongside me on this journey giving me support, and making me happy every day.To my advisor Lucas Wanner, for all the mentoring and guidance that were fundamental in formulating research questions and in my professional development.To Isaías Bittencourt Felzmann, for the feedback, ideas, collaborations on manuscripts, and the several discussions about the topics of this research.To Rodolfo Azevedo, for the support that improves the contributions of this work.

show abstract

Approximate Computing: Concepts, Architectures, Challenges, Applications, and Future Directions

Dalloo,

Jaleel Humaidi,

Al Mhdawi

et al. 2024

IEEE Access

View full text Add to dashboard Cite

SmartApprox: Learning-based configuration of approximate memories for energy-efficient execution

Cited by 3 publications

References 4 publications

EnforceSNN: Enabling resilient and energy-efficient spiking neural network inference considering approximate DRAMs for embedded systems

EnforceSNN: Enabling resilient and energy-efficient spiking neural network inference considering approximate DRAMs for embedded systems

Software and hardware interfaces for approximate memories

Approximate Computing: Concepts, Architectures, Challenges, Applications, and Future Directions

Contact Info

Product

Resources

About