Mingdai Yang
X. Sharon Hu
Abstract
Optical deep learning (DL) accelerators have attracted significant interests due to their latency and power advantages. In this article, we focus on incoherent optical designs. A significant challenge is that there is no known solution to perform single-wavelength accumulation (a key operation required for DL workloads) using incoherent optical signals efficiently. Therefore, we devise a hybrid approach, where accumulation is done in the electrical domain, and multiplication is performed in the optical domain. The key technology enabler of our design is the transistor laser, which performs electrical-to-optical and optical-to-electrical conversions efficiently. Through detailed design and evaluation of our design, along with a comprehensive benchmarking study against state-of-the-art RRAM-based designs, we derive the following key results:
(1) For a four-layer multilayer perceptron network, our design achieves 115× and 17.11× improvements in latency and energy, respectively, compared to the RRAM-based design. We can take full advantage of the speed and energy benefits of the optical technology because the inference task can be entirely mapped onto our design.
(2) For a complex workload (Resnet50), weight reprogramming is needed, and intermediate results need to be stored/re-fetched to/from memories. In this case, for the same area, our design still outperforms the RRAM-based design by 15.92× in inference latency, and 8.99× in energy.
- [1] , et al. 2019. PUMA: A programmable ultra-efficient memristor-based accelerator for machine learning inference. In Proceedings of the 24th International Conference on Architectural Support for Programming Languages and Operating Systems. 715–731.Google Scholar
Digital Library
- [2] . 2011. Realization of a narrowband single wavelength microring mirror. Appl. Phys. Lett. 99, 9 (2011).Google ScholarCross Ref
- [3] . 2017. CACTI 7: New tools for interconnect exploration in innovative off-chip memories. ACM Trans. Archit. Code Optim. 14, 2 (2017).Google ScholarDigital Library
- [4] . 2018. Silicon nitride in silicon photonics. Proc. IEEE 106, 12 (
Dec. 2018), 2209–2231.Google ScholarCross Ref - [5] . 2018. Reinforcement learning in a large-scale photonic recurrent neural network. Optica 5, 6 (2018), 756–760.Google ScholarCross Ref
- [6] . 2019. Epitaxial bonding and transfer processes for large-scale heterogeneously integrated electronic-photonic circuitry. J. Electrochem. Soc. 166, D3158 (2019).Google Scholar
- [7] . 2018. Hybrid optical-electronic convolutional neural networks with optimized diffractive optics for image classification. Sci. Rep. 8, 1 (2018), 12324.Google ScholarCross Ref
- [8] . 2020. Silicon photonics codesign for deep learning. Proc. IEEE 108, 8 (2020), 1261–1282.
DOI: Google ScholarCross Ref - [9] . 2019. Convolutional neural networks based on RRAM devices for image recognition and online learning tasks. IEEE Trans. Electron Dev. 66, 1 (2019), 793–801.
DOI: Google ScholarCross Ref - [10] . 2017. Resonance-free optical response of a vertical cavity transistor laser. Appl. Phys. Lett. 111, 12 (2017).Google ScholarCross Ref
- [11] . 2019. Large-scale optical neural networks based on photoelectric multiplication. Phys. Rev. X 9, 2 (2019), 021032.Google ScholarCross Ref
- [12] . 2018. Training of photonic neural networks through in situ backpropagation and gradient measurement. Optica 5, 7 (2018), 864–871.Google ScholarCross Ref
- [13] . 2019. Direct-modulated optical networks for interposer systems. In Proceedings of the 13th IEEE/ACM International Symposium on Networks-on-Chip.Google ScholarDigital Library
- [14] . 2020. Baldur: A power-efficient and scalable network using all-optical switches. In Proceedings of the 2020 IEEE International Symposium on High Performance Computer Architecture (HPCA’20). 153–166.Google ScholarCross Ref
- [15] . 2017. In-datacenter performance analysis of a tensor processing unit. In Proceedings of the 2017 ACM/IEEE 44th Annual International Symposium on Computer Architecture (ISCA’17).
IEEE , 1–12.Google ScholarDigital Library - [16] . 2014. A 0.65 mW/Gbps 30 Gbps capacitive coupled 10 mm serial link in 2.5D silicon interposer. In Proceedings of the 2014 IEEE 23rd Conference on Electrical Performance of Electronic Packaging and Systems. 131–134.
DOI: Google ScholarCross Ref - [17] . 2015. Electro-optical phase-change 2 × 2 switching using three- and four-waveguide directional couplers. Appl. Opt. 54, 19 (
Jul 2015 ), 5897–5902.Google ScholarCross Ref - [18] . 2018. All-optical machine learning using diffractive deep neural networks. Science 361, 6406 (2018), 1004–1008.Google ScholarCross Ref
- [19] . 2019. HolyLight: A nanophotonic accelerator for deep learning in data centers. In Proceedings of the 2019 Design, Automation & Test in Europe Conference & Exhibition (DATE’19). 1483–1488.
DOI: Google ScholarCross Ref - [20] . 2019. A mixed signal architecture for convolutional neural networks. ACM J. Emerging Technol. Comput. Syst. (JETC), Vol. 15. 1–26.Google Scholar
- [21] . 2018. PCNNA: A photonic convolutional neural network accelerator. In Proceedings of the 2018 31st IEEE International System-on-Chip Conference (SOCC’18). 169–173.
DOI: Google ScholarCross Ref - [22] . 2020. Si microring resonator crossbar arrays for deep learning accelerator. Jpn. J. Appl. Phys. 59, SG (2020), SGGE04.Google ScholarCross Ref
- [23] . 2021. Si microring resonator crossbar array for on-chip inference and training of optical neural network. arXiv:2106.04351. https://arxiv.org/abs/2106.04351Google Scholar
- [24] . 2019. A uniform modeling methodology for benchmarking DNN accelerators. In Proceedings of the 2019 IEEE/ACM International Conference on Computer-Aided Design (ICCAD’19). 1–7.Google ScholarCross Ref
- [25] . 2018. Benchmark of RRAM based architectures for dot-product computation. In 2018 IEEE Asia Pacific Conference on Circuits and Systems (APCCAS’18). 378–381.
DOI: Google ScholarCross Ref - [26] . 2020. Survey of machine learning accelerators. In Proceedings of the 2020 IEEE High Performance Extreme Computing Conference (HPEC’20). 1–12.
DOI: Google ScholarCross Ref - [27] . 1995. Adaptive multilayer optical neural network with optical thresholding. Opt. Eng. (1995), 2435–2440.Google ScholarCross Ref
- [28] . 2016. ISAAC: A convolutional neural network accelerator with in-situ analog arithmetic in crossbars. In Proceedings of the ACM/IEEE 43rd Annual International Symposium on Computer Architecture (ISCA’16).Google ScholarDigital Library
- [29] . 2017. Deep learning with coherent nanophotonic circuits. Nat. Photonics 11, 7 (2017), 441.Google ScholarCross Ref
- [30] . 2020. PIXEL: Photonic neural network accelerator. In Proceedings of the 2020 IEEE International Symposium on High Performance Computer Architecture (HPCA’20). 474–487.
DOI: Google ScholarCross Ref - [31] . 2019. CMOS-compatible, piezo-optomechanically tunable photonics for visible wavelengths and cryogenic temperatures. Opt. Express 27, 20 (
Sept. 2019), 28588–28605.Google ScholarCross Ref - [32] . 2021. CrossLight: A cross-layer optimized silicon photonic neural network accelerator. In Proceedings of the 2021 58th ACM/IEEE Design Automation Conference (DAC’21). 1069–1074.
DOI: Google ScholarDigital Library - [33] . 2021. ROBIN: A robust optical binary neural network accelerator. ACM Trans. Embedded Comput. Syst. (TECS) 20, 5s (2021), 1–24.Google ScholarDigital Library
- [34] . 2021. A survey on silicon photonics for deep learning. J. Emerg. Technol. Comput. Syst. 17, 4 (
June 2021), Article61 , 57 pages.DOI: Google ScholarDigital Library - [35] . 2017. Efficient processing of deep neural networks: A tutorial and survey. Proc. IEEE (2017), 2295–2329.Google ScholarCross Ref
- [36] . 2014. Broadcast and weight: An integrated network for scalable photonic spike processing. J. Lightwave Technol. 32, 21 (2014), 4029–4041.
DOI: Google ScholarCross Ref - [37] . 2016. Microring weight banks. IEEE J. Sel. Top. Quantum Electron. 22, 6 (2016), 312–325.
DOI: Google ScholarCross Ref - [38] . 2013. The transistor laser: Theory and experiment. Proc. IEEE 101, 10 (2013), 2271–2298.Google ScholarCross Ref
- [39] . 2021. Optical convolutional neural network with WDM-based optical patching and microring weighting banks. IEEE Photonics Technol. Lett. 33, 2 (2021), 89–92.
DOI: Google ScholarCross Ref - [40] . 2017. Investigation of statistical retention of filamentary analog RRAM for neuromophic computing. In 2017 IEEE International Electron Devices Meeting (IEDM’17). 39.4.1–39.4.4.Google Scholar
Index Terms
- A Hybrid Optical-Electrical Analog Deep Learning Accelerator Using Incoherent Optical Signals
Recommendations
A Hybrid Optical-Electrical Analog Deep Learning Accelerator Using Incoherent Optical Signals
GLSVLSI '21: Proceedings of the 2021 on Great Lakes Symposium on VLSIWe present a hybrid optical-electrical analog deep learning (DL) accelerator, the first work to use incoherent optical signals for DL workloads. Incoherent optical designs are more attractive than coherent ones as the former can be more easily realized ...
Optical information processing: A historical overview
AbstractOptical information processing lies at the intersection of optics and signal processing. It involves the processing of optical information as well as the use of optical means to process information, the later being the main emphasis of ...
Hybrid analog/digital wavelength-time optical CDMA systems in radio-over-fiber transmissions
This study presents a hybrid analog/digital (A/D) mechanism of two-dimensional (2-D) structure of wavelength-hopping and time-spreading coding optical code-division multiple-access (OCDMA) system for future generation communication and network ...
Comments