Computing enabled by electronics has been improved so extensively as to make machine learning algorithms such as deep neural network so powerful than ever before. Data communications enabled by optics have been one of the cornerstones of the modern society built upon Internet. The rapidly increasing demands for communication bandwidth due to numerous emerging applications such as AI-based cloud computing have significantly increased the bisection bandwidth of intra-datacenter networks. This trend will necessitate the optics-electronics co-packaging on board, which can only be realized by the substantial development of integrated photonics such as silicon photonics. This opportunity, making optics and electronics so close to each other, will in turn offer a chance to reconsider building EO-hybrid computational and intelligent systems. Although optical computing and optical neural networks have been proposed since a while ago, recent demonstrations of deep neural networks implemented on silicon photonics reactivate the study of exploiting photonics for machine learning. The current program-based deep neural networks are powerful, but consume huge computational resources. Suppose that just optical propagation in compact photonic chips could realize similar functions, it would greatly decrease the power and latency. Even though the scalability and the integration of nonlinear activation functions on photonic chips are still challenging, for some applications such as classifier and digit recognition, photonics systems could be viable and beneficial from such aspects.
This talk will introduce our efforts to develop the topology concepts, algorithms, and applications in order to implement silicon photonics for calculation and machine learning applications. First, a high-bit reconfigurable DAC based on generic photonic circuits will be presented. Second, without using any nonlinear activation functions or building deep neural networks, we demonstrate a photonic classifier based on only linear optical components. At last, we will show several ways of implementing silicon photonic circuit to recognize digits.
Optical devices based on silicon photonic technology are very important as supporting current and future communication
systems in terms of their scale of integration and productivity. Although various devices have been proposed and realized
using silicon photonics technology, this paper describes an optical path switch combining a large number of 2×2-unit-switches
loaded with electrical heaters on a Mach-Zehnder interferometer (MZI). We have been developed 8×8 and
32×32 optical matrix switches, which employ a path-independent-insertion-loss (PILOSS) configuration that has a
feature of the same device count on the any path of the connection. The PILOSS structure has a feature of Strictly-nonblocking,
which can connect any input port to any output port without changing existing connections when making a
Each heater of the optical switch is driven by pulse-width-modulation (PWM) generated by a Field Programmable Gate
Array (FPGA). The calibration of the pulse width according to the 2×2-unit switch state (Cross or Bar) is performed on
all of the MZIs in advance, and the values are stored in a table of the FPGA. Separately, Cross / Bar state tables
corresponding to the connection pattern of the optical input port and output port are prepared, and the pulse width
switching according to the state table is simultaneously performed based on a switching command from the upper layer
It takes a certain amount of time to change the heater temperature of the MZI arm for switching. However, it is possible
to shorten the switching time by applying a signal named “Turbo pulse” for a short term at the transition. When the
temperature is raised by the heater, the switching can be speeded-up by applying a continuous high-level voltage to the
heater temporarily. Even in the case of the temperature drops, the switching time can be accelerated by applying a
continuous high-level voltage to the heater on the other arm of the MZI. The switching time, which was actually 30μs
without Turbo pulse, could be reduced to 2.5 μs with Turbo pulse.
The fabricated silicon photonics switch chip was mounted to the chip-carrier, assembled on a printed circuit board, and
housed in a 19-inch wide 1-rack unit (RU) height blade. In addition to measuring the various characteristics of the device
in our laboratory, it has also been installed and operated at the telecommunication carrier's collocation space to confirm
long-term operation in the field. This shows that the technology related to the large-scale silicon photonics devices
shown here can be adopted to practical use.
To overcome the energy and capacity constraints in telecom and datacom networks, introduction of energy-efficient optical circuit switching (OCS) to the conventional electrical packet switching network is widely considered. In the OCS, large-port-count optical switches are essential. We have been working on optical switches based on CMOS-compatible silicon photonics that offer fast switching, compactness, low power consumption, and low-cost. Using our 45-nm CMOS process, we have recently demonstrated a low-loss 32 x 32 silicon photonics switch, and their performance improvements of wide operation bandwidth, polarization-independent operation, etc. We review the recent progress of our silicon photonics switches.
This paper discusses how to realize an optical circuit switching interconnect capable of more than 10 Tbps link bandwidth and more than 100,000 end points scalability. To keep continuous performance improvement of datacenters or high performance computers, high capacity and low latency interconnect network is essential. To handle such large bandwidth interconnect networks with low energy consumption, optical switch technologies will become inevitable. This paper firstly examines the scaling of the energy consumption of optical circuit switching networks based on the state of the art silicon photonics switch technology. Secondly to achieve Tbps-class link bandwidth, the WDM transmission technology and a shared WDM light source mechanism named “wavelength bank” are introduced. Due to the shared light source, each optical transceiver does not have to carry individual light sources, which enables simple WDM transceivers with cost-efficient silicon photonics technologies. Then a new optical switch control approach which reduces the control overhead time is discussed. In the proposed approach, the optical data plane itself represents the path destination, which enables a simple distributed-like control procedure. The proposed approach is expected to achieve the scalability and flexibility supporting more than 10 Tbps link bandwidth and more than 100,000 endpoints with 40 WDM channels. The proposed interconnect architecture offers direct end-to-end optical paths enabling low latencies with the “speed of light”. The paper also discusses some of the challenges which should be resolved to practically realize the future large bandwidth optical interconnect networks.
Intra-datacenter traffic is growing more than 20% a year. In typical datacenters, many racks/pods including servers are interconnected via multi-tier electrical switches. The electrical switches necessitate power-consuming optical-to- electrical (OE) and electrical-to-optical (EO) conversion, the power consumption of which increases with traffic. To overcome this problem, optical switches that eliminate costly OE and EO conversion and enable low power consumption switching are being investigated. There are two major requirements for the optical switch. First, it must have a high port count to construct reduced tier intra-datacenter networks. Second, switching speed must be short enough that most of the traffic load can be offloaded from electrical switches. Among various optical switches, we focus on those based on arrayed-waveguide gratings (AWGs), since the AWG is a passive device with minimal power consumption. We previously proposed a high-port-count optical switch architecture that utilizes tunable lasers, route-and-combine switches, and wavelength-routing switches comprised of couplers, erbium-doped fiber amplifiers (EDFAs), and AWGs. We employed conventional external cavity lasers whose wavelength-tuning speed was slower than 100 ms. In this paper, we demonstrate a large-scale optical switch that offers fast wavelength routing. We construct a 720×720 optical switch using recently developed lasers whose wavelength-tuning period is below 460 μs. We evaluate the switching time via bit-error-ratio measurements and achieve 470-μs switching time (includes 10-μs guard time to handle EDFA surge). To best of our knowledge, this is the first demonstration of such a large-scale optical switch with practical switching
Dynamic path switching in lower layers such as optical or sub-wavelength layer-1 path connections is essential for future
networks to provide end-to-end, bandwidth-guaranteed, large-capacity services without energy crunch. While this is
almost generally agreed, the number of ports in optical switches tends to be limited by technological difficulties, severely
restraining the scale of the network. However, video-related services, that occupies most of the traffic nowadays, could
significantly alleviate such restraints if we utilized the nature of video usage. In most cases, video-related services are
virtually provided through prior reservation scheme in which a relatively high call-blocking probabilities or long latency
for a connection can be tolerated. This situation allows us to accommodate a relatively high number of subscribers with a
limited number of switch ports.
This paper shows that a network using optical switches with a technologically feasible number of ports, multi-granular
paths, and a hierarchical network topology can be of a national scale accommodating several tens of millions of
subscribers. The purpose of detailing a plausible network topology is to show that such a network offers a benefit of
energy efficiency approximately three orders of magnitude compared with that extrapolated from recent router-based
We then discuss important technical aspects of such dynamic optical path networks including our several research
activities. We emphasize the importance of vertically integrated research activities from application to device layers to
develop the dynamic optical path networks.
We introduce parametric tunable dispersion compensation scheme that has strong advantages in operating bandwidth and
tuning response. The parametric tunable dispersion compensator (P-TDC) using highly-nonlinear fiber with a low
dispersion slope achieves wide grid-less operating bandwidth of more than 1 THz and fast tuning responses of a few tens
of microseconds. We also develop the setup of the P-TDC for practical usage, utilizing polarization diversity scheme.
Field transmission using the polarization-insensitive P-TDC is also successfully demonstrated.
This invited talk will review the development of ultrafast all-optical LAN technologies, conducted by New Energy and
Industrial Technology Development Organization (NEDO), Japan. First, we will provide an outlook for the energy issues
of future network equipment, then point out the importance of optical circuit-switched networks, particularly for the
future local area networks in the forthcoming ultra-high definition, or 'Super Hi-vision', video era. To realize ultrafast
all-optical LAN, we argue that scalable network interface card technologies are the key. As specific development topics,
40G-CMOS based optical transceivers, picosecond all-optical switching using the inter-subband transition (ISBT)
devices, high-operating-temperature semiconductor optical amplifiers (SOA), and integrated wide-dynamic-range
wavelength converters will be introduced.
We experimentally investigate the generation of a low-noise ultra short pulse train from 40GHz to160GHz by using Comb-like profiled fiber (CPF) for adiabatic soliton conversion and compression. Highly nonlinear fibers allow us to reduce total length of CPF as well as to utilize Kerr effect in the fiber effectively. We demonstrate generations of 160GHz soliton train of 750fs, the compression to 500fs of 40GHz externally-modulated pulse with wideband tunability over 30nm. Then we apply the CPF pulse compression technique to achieve the programmable repetition tunability from 5 to 500 MHz in low pedestral 300fs pulse train generation.
R and D activities on photonic networks in Japan are presented. First, milestones in current, ongoing R and D programs supported by Japanese government agencies are introduced, including long-distance and WDM fiber transmission, wavelength routing, optical burst switching, and control plane technology for IP backbone networks. Their goal was set to evolve a legacy telecommunications network to IP over WDM networks by introducing technologies for WDM and wavelength routing. We then discuss the perspectives of so-called PHASE II R and D programs for photonic networks over the next five years until 2010, by focusing on the report which has been recently issued by the Photonic Internet Forum (PIF), a consortium that has major carriers, telecom vendors, and Japanese academics as members. The PHASE II R and D programs should serve to establish a photonic platform to provide abundant bandwidth on demand, at any time on a real-time basis through the customer's initiative, to promote bandwidth-rich applications, such as grid computing, real-time digital-cinema streaming, medical and educational applications, and network storage in e-commerce.
Recent development on Raman amplifiers for WDM communication is reviewed. The design and demonstration of Raman amplifiers are summarized with concentration on Raman gain equalization by using multi-wavelength pumping scheme. Polarization dependence of Raman gain is measured against the degree of polarization of pump source. In order to realize stable and low degree of polarization, it is investigated how to design a passive depolarizer for pump laser diodes. Finally, a grating-integrated pump laser diode is developed in order to reduce relative intensity noise, which can cause severe degradation of amplifier characteristics.
Optical amplifiers are demanded to improve their performances in order to support the forthcoming Internet oriented WDM networks. This paper reviews recent topics on next generation optical fiber amplifier technologies developed at author's research group.