Albireo: Energy-Efficient Acceleration of Convolutional Neural Networks via Silicon Photonics

Shiflett, Kyle; Kodi, Avinash; Bunescu, Răzvan; Louri, Ahmed

doi:10.1109/isca52012.2021.00072

Cited by 25 publications

(16 citation statements)

References 52 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In addition to being the photo-to-electric conversion devices in optical neural networks, photodetectors are also used to realize the operations of ANNs. In [59] the positive signals and negative signals are propagated and superimposed on different optical waveguides, and two balanced photodetectors are set at the end of the waveguides to detect the total optical power in positive and negative waveguides, respectively. The detected lights are converted into the currents by the photodetectors.…”

Section: Photonic Devicesmentioning

confidence: 99%

Efficient neural network accelerators with optical computing and communication

et al. 2023

View full text Add to dashboard Cite

Conventional electronic Artificial Neural Networks (ANNs) accelerators focus on architecture design and numerical computation optimization to improve the training efficiency. However, these approaches have recently encountered bottlenecks in terms of energy efficiency and computing performance, which leads to an increase interest in photonic accelerator. Photonic architectures with low energy consumption, high transmission speed and high bandwidth have been considered as an important role for generation of computing architectures. In this paper, to provide a better understanding of optical technology used in ANN acceleration, we present a comprehensive review for the efficient photonic computing and communication in ANN accelerators. The related photonic devices are investigated in terms of the application in ANNs acceleration, and a classification of existing solutions is proposed that are categorized into optical computing acceleration and optical communication acceleration according to photonic effects and photonic architectures. Moreover, we discuss the challenges for these photonic neural network acceleration approaches to highlight the most promising future research opportunities in this field.

show abstract

Section: Photonic Devicesmentioning

confidence: 99%

Efficient neural network accelerators with optical computing and communication

et al. 2023

View full text Add to dashboard Cite

show abstract

“…Prior works [31], [21], and [19] have analysed the scalability (i.e., achievable value of VDPE size N under the constraints of bit precision and data rate) of AMM and MAM VDPCs. Table I reproduces the supported values of VDPE size N (considering M=N) for AMM and MAM VDPCs at various data rates (DRs) and bit precision from [21].…”

Section: A Scalability Limitations Of Mrr-based Analog Vdpcsmentioning

confidence: 99%

“…For MAM VDPC for 1 GS/s, maximum N reduces from 44 to 12 as we increase the input/weight precision from 4-bit to 6-bit. The reason for such strong tradeoff between N and achievable input/weight precision (referred to as B, henceforth) in MAM and AMM VDPCs is that both B and N strongly depend on the number of distinguishable analog optical power levels [21] [31], which is proportional to N × 2 B . Hence, for a fixed number of distinguishable analog optical power levels, which is defined by the analog optical power resolution of the utilized summation elements (SEs) (see SEs in Fig.…”

Section: A Scalability Limitations Of Mrr-based Analog Vdpcsmentioning

confidence: 99%

“…We omitted comparison with CMOS-based digital CNN accelerators as 2.55 0.002 0.78ns Serializer per OSM [48] 5 5.9 0.03ns LUT per OSM [49] 0.06 0.09 2ns PCA [44] 0.02 0.28 prior analog optical photonic CNN accelerators have outperformed them [9], [12]. We simulate analog optical accelerators for 5 GS/s [31] and from Section III-A, at B=4-bit precision, we set N=16 for AMM (DEAPCNN), and N=22 for MAM (HOLYLIGHT). Prior works, AMM (DEAPCNN) and MAM (HOLYLIGHT) employ weight stationary dataflow, therefore our evaluation is based on weight stationary dataflow.…”

Section: B Simulation Setupmentioning

confidence: 99%

See 1 more Smart Citation

SCONNA: A Stochastic Computing Based Optical Accelerator for Ultra-Fast, Energy-Efficient Inference of Integer-Quantized CNNs

Vatsavai¹,

Karempudi²,

Thakkar³

et al. 2023

Preprint

View full text Add to dashboard Cite

Convolutional Neural Networks (CNNs) are used extensively for artificial intelligence applications due to their record-breaking accuracy. For efficient and swift hardware-based acceleration, CNNs are typically quantized to have integer input/weight parameters. The acceleration of a CNN inference task uses convolution operations that are typically transformed into vector-dot-product (VDP) operations. Several photonic microring resonators (MRRs) based hardware architectures have been proposed to accelerate integer-quantized CNNs with remarkably higher throughput and energy efficiency compared to their electronic counterparts. However, the existing photonic MRRbased analog accelerators exhibit a very strong trade-off between the achievable input/weight precision and VDP operation size, which severely restricts their achievable VDP operation size for the quantized input/weight precision of 4 bits and higher. The restricted VDP operation size ultimately suppresses computing throughput to severely diminish the achievable performance benefits. To address this shortcoming, we for the first time present a merger of stochastic computing and MRR-based CNN accelerators. To leverage the innate precision flexibility of stochastic computing, we invent an MRR-based optical stochastic multiplier (OSM). We employ multiple OSMs in a cascaded manner using dense wavelength division multiplexing, to forge a novel Stochastic Computing based Optical Neural Network Accelerator (SCONNA). SCONNA achieves significantly high throughput and energy efficiency for accelerating inferences of high-precision quantized CNNs. Our evaluation for the inference of four modern CNNs at 8-bit input/weight precision indicates that SCONNA provides improvements of up to 66.5×, 90×, and 91× in frames-per-second (FPS), FPS/W and FPS/W/mm 2 , respectively, on average over two photonic MRR-based analog CNN accelerators from prior work, with Top-1 accuracy drop of only up to 0.4% for large CNNs and up to 1.5% for small CNNs. We developed a transaction-level, event-driven python-based simulator for the evaluation of SCONNA and other accelerators (https://github.com/uky-UCAT/SC ONN SIM.git).

show abstract

“…As far as we know, existing optical computing system does not consider the trade-off between the energy consumption of ADCs and the accuracy of optical computing results. For example, SONIC, Albireo and DEAP [3,4,5]use 16-bit, 8-bit and 7-bit ADCs respectively, but they only perform well in simple DNNs, such as LeNet. For LightBulb and Holylight [7,8], 4-bit and 8-bit ADCs are used respectively, but the models are not common.…”

Section: Introductionmentioning

confidence: 99%

An energy-efficient quantized inference framework for electro-photonic computing system

Lin

Zhang

et al. 2023

International Conference on Optical and Photonic Engineering (icOPEN 2022)

View full text Add to dashboard Cite

Deep Learning is developing rapidly and has made breakthroughs in various fields. As a result, the number of parameters in deep neural networks(DNNs) is scaling about 5× rate of Moore's Law. To meet the computing power needs of DNNs, new computing architectures have been proposed, and the electro-photonic computing system is one of them. With the physical characteristics of photons, the electro-photonic computing system has demonstrated great potential in improving the efficiency of convolution calculation. However, there are still much spcace for improvement in the energy efficiency of the electro-photonic computing system. And the high energy consumption led by the high-precision analogto-digital converters(ADCs) is an important factor hindering the energy efficiency. ADCs dominate about 50% energy consumption of the entire system, and they will increase exponentially with the precision. But if the low-precision ADCs are used directly, they will destroy the computing accuracy in the convolution calculation and reduce the inference performance of DNNs. In this paper, we propose an energy-efficient quantized inference framework for electro-photonic computing system to balance the energy consumption of ADCs and the DNNs inference performance, which includes a multi-scale quantization scheme based on per-slice granularity and width variable optical matrix. The proposed framework can reduce the energy consumption of the system by using low-precision ADCs, and ensure the inference performance of the DNNs at the same time. The experiments are carried out in MobileNet-v2, ResNet50 and Vgg16 respectively. The experimental results show that the proposed quantized inference framework can reduce the precision of ADCs by 7 bits and compared with using the low-precision ADCs directly, it improves the model accuracy of 40.5%, 70.2% and 65.4% respectively , which is close to the full precision model.

show abstract

Albireo: Energy-Efficient Acceleration of Convolutional Neural Networks via Silicon Photonics

Cited by 25 publications

References 52 publications

Efficient neural network accelerators with optical computing and communication

Efficient neural network accelerators with optical computing and communication

SCONNA: A Stochastic Computing Based Optical Accelerator for Ultra-Fast, Energy-Efficient Inference of Integer-Quantized CNNs

An energy-efficient quantized inference framework for electro-photonic computing system

Contact Info

Product

Resources

About