Overview
Over the past five years, sensor sensitivity and accuracy have improved by an order of magnitude while power consumption, cost, and size have fallen by about 80%. Over the next decade, sensors will need to become more complex, more reliable, lower cost, and offer high-bandwidth interconnects to support broader deployment in wearables, artificial intelligence, autonomous driving, and other areas.
Integrated Photonics, CPO and PIC
As transistor scaling in two dimensions slows and 2.5D/3D packaging matures, integrated photonics manufactured in CMOS fabs and packaged with advanced integrated circuit packaging technologies is becoming a key innovation area. Optical transceivers and interconnect devices co-located in the same package with data processing and memory chips are referred to as co-packaged optical devices (CPO). Chips that contain active photonic devices and associated photonic circuits are called photonic integrated circuits (PIC). These PICs convert electrical signals to optical signals and carry optical-domain data between packages or between compute cores and memory inside packages.
The primary value proposition of CPO is increased bandwidth density and energy efficiency, driven by growing compute capacity and communication bandwidth demand. This demand is especially influenced by the rapid growth of complex AI and ML accelerators and compute clusters, which push interconnect bandwidth, energy efficiency, and lower latency from chip-scale systems to large systems composed of thousands of GPUs, CPUs, and memory ICs.
Sensors and Actuators in Optical Systems
As co-packaged optical transceivers improve data transport efficiency, optical sensors and actuators play an increasingly important role in data collection and processing. These sensors and actuators are critical in products such as displays, automotive headlights, projectors, DNA analysis chips, and optical switches for data sensors and wearable heart-rate and oxygen sensors. Many of these devices are manufactured using MEMS technology. MEMS devices and fabrication techniques are also used to align, modulate, or tune optical devices such as tunable filters, lasers, and fibers, enabling products like near-infrared (NIR) material-analysis scanners.
The combination of MEMS and application-specific integrated circuits (ASICs) is maturing, enabling novel products. Sensor miniaturization continues to drive telemedicine, smart homes, smart cities, advanced manufacturing, and handheld wearables, all of which demand sensors that are more capable, reliable, lower cost, and offer high-bandwidth interconnects.
Sensor Trends for the Next Decade
Handset economics are driven by cost, size, performance, and bandwidth. Robust sensor design is critical for GPS, gyroscopes, accelerometers, pressure sensors, magnetometers, optical image stabilization, microphones, and fingerprint sensing. Sensor sensitivity and accuracy have improved by 10x in the last five years while power, cost, and size have fallen significantly. These trends are expected to continue.
The integration of physical sensing and AI at the device level has improved MEMS-based product capability. Integrated sensors enable seamless navigation, stability control, impact detection, adaptive lighting, image stabilization, and traction control. Better sensor performance yields higher signal-to-noise ratio, larger dynamic range, and sub-milliwatt power consumption.
Additional sensor capabilities will be required. For example, silicon-integrated elements with narrower bandgap than silicon that operate in the short-wave infrared (SWIR) are needed for high-resolution, eye-safe long-range LiDAR. Packaging these sensors in ultra-compact and flexible packages is desirable for wearables and medical applications.
Edge computation and intelligent processing near sensors are critical for energy efficiency and latency, with co-optimization of hardware and software being a key enabler. TinyML (tiny machine learning) is a rapidly developing area for sensors and actuators, including hardware, algorithms, and software deployed on or near edge devices. Typically, ML inference on sensor data is performed locally with power budgets on the order of 1 mW.
As TinyML proliferates in the IoT, focus areas will include low-power wake circuits, integration of nonvolatile memories manufactured in advanced silicon nodes, and ML algorithms that use limited memory and compute within milliwatt power budgets.
Silicon photonics promises to extend frequency and bandwidth for sensing, interconnects, communications, and computation. Advanced optical sensors for health and medical sensing, including glucose monitoring, blood pressure, and cardiac markers, as well as automotive sensors like LiDAR, present significant opportunities. A key challenge for these sensors is meeting the accuracy and reliability of incumbent solutions.
True solid-state beam steering for ADAS LiDAR requires further development of optical phased arrays and gratings. Solid-state beam steering will lower cost, improve reliability, and shrink LiDAR systems, all of which are needed to enable mass production of SAE Level 3 and above autonomous vehicles.
Systems that consume sensor data often trust the sensor output without additional validation. Manipulating the physical phenomena that sensors are designed to interpret can therefore produce undesired behaviors. For example, smartphone MEMS accelerometers can be tricked into registering steps by playing barely audible sounds from a video. Such techniques could also be used to transmit information covertly.
Consequently, MEMS device security is a challenge that must be addressed to prevent malicious tampering with sensor data. MEMS can also be part of security solutions, serving as physical protection mechanisms that make it harder to tamper with circuits.
Key Technology and Product Trends
- Replace crystal oscillators with MEMS-based resonators compatible with CMOS to enable new architectures, higher performance, and removal of off-chip passive components.
- Use piezoelectric MEMS sensors and actuators for handheld ultrasound and improve and miniaturize speakers and microphones.
- Apply a range of technologies and low-cost materials to mass-produce lower-cost sensors while still manufacturing high-precision sensors for critical tasks such as GNSS navigation.
- Embed sensors into garments and fabrics to create new product categories that compete with phone-, ring-, patch-, and watch-based sensors. Fabric sensors introduce new requirements for connectivity, reliability, and durability.
- MEMS structures are being used in quantum computing to enable qubit interfaces with the external world.
- Sensor manufacturing processes are diverse and frequently application-specific. Some sensors need to be open to the environment while resisting unwanted influences; others are better suited to sealed packaging. Flexible substrates combining flexible and rigid sensors are an important packaging direction, as is continued movement toward standardized sensor package solutions to improve supply-chain efficiency.
Integration Strategies
Sensors can be integrated with electronics in multiple ways: fabricated on the same chip using compatible processes, created as post-CMOS layers above or below CMOS, or assembled as discrete chips. MEMS can be detached from the CMOS stack, and some researchers use fin structures as sensing elements. Post-processing and separated chips can leverage advanced CMOS nodes and optimized MEMS processes. Each integration strategy affects packaging, material selection, manufacturing, and assembly, and each will have advantages in the foreseeable future.
Challenges remain in efficient electro-optical conversion and interface integration with respect to power and area. Addressing these challenges is necessary to expand the application space of photonics.
Communications Trends
Integrated photonics will greatly improve communications infrastructure. In the next five years, symbol rates are expected to exceed 100 Gbaud, and aggregated bandwidth of integrated photonic transceiver modules for fiber communications could reach 3.2 to 6.4 Tbit/s. The urgent need for higher bandwidth and energy efficiency will drive data-center architectures from pluggable optics toward co-packaged optics (CPO), accelerating the shift from copper to short-reach optical fiber below one meter.
Dense wavelength-division multiplexing (DWDM) or similar multiplexing is required to fully exploit fiber bandwidth. Implementing DWDM or coherent techniques on ASICs enables tradeoffs between per-channel data rate and channel count for a given fiber bandwidth, improving energy efficiency and lowering system cost. As photonic interconnect bandwidth density increases, reducing system energy per bit, cost per bit, and latency will be increasingly important, especially for AI and mobile computing. Packaging optical devices close to ASICs and other compute ICs using short, low-loss channels helps improve energy efficiency and bandwidth, enabling low-power electro-optical conversion and high-bandwidth data transport at large scale.
Analog photonic links are being used to simplify millimeter-wave node IC architectures, enabling over 1000 antennas per chip in some cases. Analog photonic links can also achieve energy efficiencies in the signal chain that are far better than digital links.
Large portions of the electromagnetic spectrum remain underutilized. Unlocking this spectrum will require semiconductor innovations such as SOI/SiGe photonics, VCSELs, micro-LEDs, avalanche photodiodes, and InP-based PICs to provide advanced process platforms for future bandwidth and power improvements.
In MEMS, optical switches are replacing OEO switches. MEMS-based tunable filters, capacitors, RF switches, and MEMS resonators are enabling new architectures and higher integration, reducing footprint and simplifying packaging. MEMS-enabled tunability, alignment, and calibration increase specifications and performance for communications components.
Computation and Memory
Photonics offers a path to relieve power and bandwidth bottlenecks caused by moving large amounts of data in high-performance and data-center computing. Key integrated photonic components include high-speed, low-power transceivers, light sources, waveguides/modulators, and photodetectors. A continuing challenge is scalable manufacturing and sustainable size scalability for these devices. Photonics provides opportunities for low-power, high-speed I/O and photonic interconnect topologies. Photonic devices have also been demonstrated for certain mathematical operations, such as matrix-vector multiplication.
Integrating photonics near processors and on links from processors to memory can exploit the large bandwidth and low-loss advantages of optical transmission compared with copper, potentially increasing bandwidth density by more than 100x and improving energy efficiency by more than 10x at the package level.
Optical links enable low-energy, low-latency interconnects that support disaggregation of networking, compute, and memory. This approach leverages optical devices co-packaged with ASICs/CPUs/GPUs and optical communications standards to connect cross-rack compute and memory in hyperscale data centers.
AI accelerators and neuromorphic computing are additional areas that, when combined with advanced 2.5D/3D packaging, have significant potential to improve energy efficiency.
MEMS-based products require more on-board processing to produce smarter sensors. Typical MEMS sensors today provide processed outputs such as intelligent motion detection. This reflects the trend toward intelligent edge processing and is reshaping CMOS-MEMS integration and advanced packaging.
New Materials and Processes
Innovative semiconductor process platforms include SOI/Ge photonics; heterogeneous integration of III-V materials on silicon wafers via epitaxial growth, wafer or die bonding, or laser cavity coupling for light sources, modulators, and detectors; and active devices based on physical effects beyond plasma dispersion, such as plasmonics and graphene.
Thin films of lithium niobate and barium titanate enable high-frequency modulation for hybrid integration, while photonic wire bonds using 3D printing for laser packaging and interconnects require further exploration. High-bandwidth, low-loss, low-power chip-to-chip optical interconnects will likely need advances in substrate and PCB-embedded waveguides. It is important to ensure low-cost, reliable photonic materials that remain stable over long durations and wide temperature ranges while maintaining low thermal hysteresis and low loss to minimize overall energy consumption.
For certain sensors and actuators, CMOS-compatible extensions are needed to improve inertial sensor performance and RF filter power handling, prompting exploration of new materials such as tungsten or other high atomic weight metals to reduce overall footprint. Aluminum nitride CMOS-compatible platforms are also under investigation, as are new piezoelectric and phase-change memory sensors with higher force density and linearity. These new materials for sensors and actuators in turn drive innovations in packaging materials and techniques that control temperature and humidity.
Two important material drivers for inertial sensors are preventing mechanical structures from stiction to the substrate and selecting material groups with compatible coefficients of thermal expansion to avoid unwanted bending and stress, especially stress induced by packaging. These sensors also need improved affordable sourcing of materials, coatings that resist wear, stiction, corrosion, and charge build-up. Chemical sensors require materials that promote chemical reactions in repeatable and stable ways. Acoustic sensor and actuator materials, such as those used for microphones and speakers, must enable controlled motion to generate, move, and detect sound.
Because these devices endure extensive cycling, a key characteristic of MEMS materials is predictable deformation without fatigue. Optical sensor and actuator materials require flat, reflective surfaces that resist deformation and have appropriate optical properties. For RF MEMS switches, developing reliable contact materials remains critical.
Technical Status and Product Examples
Efforts to introduce photonics into integrated circuit packaging have begun. In recent years, initial products transitioning from standalone transceivers to CPO or optical engines designed for CPO have been demonstrated.
Advanced MEMS products combine multiple sensors with electronics to provide processed outputs via on-board low-power electronics, often integrating AI and advanced calibration. Emerging MEMS products blend advances in sensor/actuator fabrication, materials, and design. Examples include:
- MEMS-based speakers
- Chemical sensors capable of detecting multiple substances
- Ultrasound arrays for handheld imaging instruments
- BAW devices that can integrate with CMOS
Figure 5, for example, describes an advanced MEMS product that combines proven piezoelectric imaging performance with silicon cost advantages in a low-voltage pMUT array chip. Each chip contains 4,096 independently controlled pMUTs, offering large bandwidth, high sensitivity, and up to 150-degree field of view for rapid imaging improvements and real-time AI frame analysis. Looking ahead, pMUT technology may enable robust 3D imaging and potentially 4D imaging for faster diagnosis in medical care.
Limitations of Current Technologies
Optical communications are not new. Decades ago, the maturation of fiber manufacturing and the advent of III-V-based light emitters and detectors expanded the role of optical transceivers in long-distance data transmission due to fiber's low loss, low dispersion, and wide bandwidth. Advances and diversification of technology reduced costs and enabled widespread use of optical transceivers in access and client applications.
The emergence of DWDM and erbium-doped fiber amplifiers (EDFAs) transformed long-haul networks and established the foundation for modern high-speed communication. Optical transceivers evolved from custom devices to on-board modules and to standardized small form-factor pluggables that are ubiquitous in today's data centers.
However, as data volumes and compute capability grow, standalone transceivers are becoming a bandwidth bottleneck. The number of pluggable transceivers that can be mounted on panels may be insufficient to provide the bandwidth required by line cards or data switches. At symbol rates of 100 Gbaud and beyond, copper link losses between IC serializers and transceiver inputs present signal integrity challenges, making CPO an attractive alternative to break the physical barrier while continuing to reduce system power and cost.
Placing optical transceivers in the same package resolves signal integrity issues by enabling much shorter, lower-loss interconnects and potentially direct drive of optical devices. To overcome bandwidth density limits, however, DWDM techniques for longer-reach transceivers may still be needed while preserving short-reach co-packaged optics for local connections.
Because MEMS devices often require custom processes, creating highly standardized platforms like CMOS is difficult. Different sensor types can demand different materials. Some manufacturers attempt multi-sensor fabrication processes on the same chip, but cost and performance considerations typically favor custom processes. Design rules must therefore encode MEMS fabrication and associated packaging considerations.
Another limitation is the lack of primitive elements for MEMS comparable to CMOS transistors. Testing methods are often specific to a sensor's working principle or application, making test and assembly infrastructure, ecosystems, and supply chains more complex than for electronic products. While progress has been made, these constraints still limit the widespread adoption of MEMS and can delay time to market. Continued market growth should encourage more suppliers to enter the MEMS space and improve these aspects.
Challenges and Potential Solutions
Communications, compute, and memory applications face technical and supply-chain challenges, including:
- PIC performance (200G per lambda and above), yield, manufacturability, and cost
- Low-cost DWDM III-V laser sources for silicon photonics at O and C bands suitable for mass production
- DFB, QD, and other laser sourcing options with target performance, wall-plug efficiency, and cost
- Non-hermetic lasers that perform reliably at elevated temperatures in IC package environments
- Integration of high-power lasers with high yield and reliability
- Heterogeneous integration of laser materials
- Edge- and vertical-coupling fiber attachment schemes, fiber pitch scaling, and cost-effective solutions
- High-throughput, low-loss fiber assembly processes
- Ribbonization of fiber arrays with higher fiber counts and reduced cladding diameters
- Development and standardization of fiber-array interconnect/termination hardware
- Fiber management for CPO with high fiber density per package and ribbonized fibers
- Advanced heterogeneous packaging, including 3D TSVs for high bandwidth density, high signal integrity, and low interconnect power
- Optical bus architectures for processor-processor and processor-memory access
- Low-cost light sources for optical interconnects, thermal tunability and junction-temperature management, and design for testability (DFT) and design for manufacturability (DFM)
- Stable, viable supply chains and ecosystems for PICs, lasers, and fiber integration
- Modeling standards for photonic circuits to support ecosystem design
Sensor and actuator applications also face multiple technical and supply-chain challenges:
- CAD: reduced-order nonlinear modeling for sensors and actuators; MEMS co-design with electronics and packaging; PDKs with material properties across relevant physical domains.
- Materials: characterization of new materials across relevant physical domains; materials synthesis tools for discovery and optimization; characterization of bending and stretching behaviors, especially for wearables.
- Standards: standards for material bending and stretching performance; sensor performance FOM standards; reliability and test standards for emerging technologies.
- CMOS and multi-sensor integration: continued transition to advanced packaging methods from stacked wire-bonded sensors to enable greater heterogeneous integration.
- Sensor design and manufacturing improvements: use of field calibration, multiple sensors, and fusion with non-MEMS sensors to elevate inertial sensors to navigation grade; improved design and manufacturing methods and broader process windows to compensate for manufacturing nonidealities; MEMS energy harvesters must increase conversion efficiency to compete with solar and thermoelectric sources; ongoing development of low-power and near-zero-power sensors to meet energy constraints; optical glucose sensors must improve accuracy to compete with needle-based electrochemical sensors; accuracy improvements are required for paper and plastic sensors to compete with silicon-based sensors; continued research into atomic clock technologies to replace large components.
