Quantifying power use in silicon photonic neural networks

Tait, Alexander N.

doi:10.48550/arxiv.2108.04819

Cited by 2 publications

(2 citation statements)

References 59 publications

(138 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…One limitation of the performance of all optoelectronic neural networks is the finite signal to noise ratio (SNR) of modern photoreceivers [25]. Loss from fiber propagation or diffraction in free space links can force the client to operate in a photon starved environment.…”

Section: Receiver Sensitivitymentioning

confidence: 99%

Delocalized Photonic Deep Learning on the Internet's Edge

Sludds

Bandyopadhyay

Chen

et al. 2022

Preprint

View full text Add to dashboard Cite

Advances in deep neural networks (DNNs) are transforming science and technology. However, the increasing computational demands of the most powerful DNNs limit deployment on low-power devices, such as smartphones and sensors – and this trend is accelerated by the simultaneous move towards Internet-of-Things (IoT) devices. Numerous efforts are underway to lower power consumption, but a fundamental bottleneck remains due to energy consumption in matrix algebra, even for analog approaches including neuromorphic, analog memory and photonic meshes. Here we introduce and demonstrate a new approach that sharply reduces energy required for matrix algebra by doing away with weight memory access on edge devices, enabling orders of magnitude energy and latency reduction. At the core of our approach is a new concept that decentralizes the DNN for delocalized, optically accelerated matrix algebra on edge devices. Using a silicon photonic smart transceiver, we demonstrate experimentally that this scheme, termed Netcast, dramatically reduces energy consumption. We demonstrate operation in a photon-starved environment with 40 aJ/multiply of optical energy for 98.8% accurate image recognition and <1 photon/multiply using single photon detectors. Furthermore, we show realistic deployment of our system, classifying images with 3 THz of bandwidth over 86 km of deployed optical fiber in a Boston-area fiber network. Our approach enables computing on a new generation of edge devices with speeds comparable to modern digital electronics and power consumption that is orders of magnitude lower.

show abstract

Section: Receiver Sensitivitymentioning

confidence: 99%

Delocalized Photonic Deep Learning on the Internet's Edge

Sludds

Bandyopadhyay

Chen

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…Computational requirements and energy expenditures have escalated rapidly, to either process the exponentially increased data generated by ultra-high-speed mobile networks or to address the demand for accelerating artificial intelligence 1 . However, current state-of-art electronic processors, which have developed with startlingly rapid progress in the past decades, are approaching their growth limit subject to Moore's Law.…”

Section: Introductionmentioning

confidence: 99%

Towards a High-density Photonic Tensor Core Enabled by Intensity-modulated Microrings and Photonic Wire Bonding

Luan

Salmani

et al. 2022

Preprint

View full text Add to dashboard Cite

We propose a photonic processing unit for high-density analog computation using intensity-modulation-based microringmodulators (IM-MRMs). The output signal at the fixed resonance wavelength is directly intensity modulated by changing theextinction ratio (ER) of the IM-MRM. Thanks to the intensity-modulated approach, the proposed photonic processing unit isless sensitive to the inter-channel crosstalk. Simulation results reveal that the proposed design offers a maximum of 17-foldincrease in wavelength channel density compared to its wavelength-modulated counterpart. Therefore, a photonic tensor coreof size 512 × 512 can be realized by current foundry lines. A convolutional neural network (CNN) simulator with 6-bit precisionis built for handwritten digit recognition task using the proposed modulator. Simulation results show an overall accuracy of 96.76%, when the wavelength channel spacing suffers a 3-dB power penalty. To experimentally validate the system, 1000 dotproduct operations are carried out with a 4-bit signed system on a co-packaged photonic chip, where optical and electrical I/Osare realized using photonic and electrical wire bonding techniques. Study of the measurement results show a mean squarederror (MSE) of 3.09 × 10^−3 for dot product calculations. The proposed IM-MRM, therefore, renders the crosstalk issue tractableand provides a solution for the development of large-scale optical information processing systems with multiple wavelengths.

show abstract

Quantifying power use in silicon photonic neural networks

Cited by 2 publications

References 59 publications

Delocalized Photonic Deep Learning on the Internet's Edge

Delocalized Photonic Deep Learning on the Internet's Edge

Towards a High-density Photonic Tensor Core Enabled by Intensity-modulated Microrings and Photonic Wire Bonding

Contact Info

Product

Resources

About