Silicon photonics is a technique used to prepare photonic integrated circuits (PICs), typically used to generate, detect, transmit, and process light. This method uses a silicon on semiconductor insulator (SOI) wafer as the substrate material and is fabricated using standard complementary metal-oxide-semiconductor (CMOS) semiconductor technology.
The growing demand for fast and efficient communication technologies has led to an increase in research in silicon photonics. Since its inception, silicon photonics has been developing rapidly and has been adopted as a powerful technology by many foundries that manufacture devices such as modulators, photodetectors, and lasers.
Compared with traditional methods, photonics-based computing consumes less energy and transmits faster dataBut the cost of manufacturing silicon integrated circuits embedded in photonic components has hindered the development of this technology
Formfactor's groundbreaking Cascade CM300Xi-Siph photonic functional system fully integrates PI's fast multi-channel photonic alignment engine to enable high-throughput, nanometer-precision optical probing of silicon photonic devices on wafers without the need for further customer development or engineering resources.
Now, the global demand for data is growing exponentially, driven by a variety of applications such as social networks and streaming**, genomics-driven medicine, and the proliferation of connected devices in the "Internet of Things."
The growth of mobile computing is particularly pronounced. According to the United Nations Telecommunications Agency, by 2013, the number of mobile** subscribers was close to the number of people on the planet. In contrast, the number of fixed** users has never exceeded 25% of the global population.
The wealth of opportunities offered by these technologies and services, and any opportunities they provide in the future, will ensure that people will continue to look for new ways to connect, entertain, inform, and help us with data. However, the increased use of data comes at a cost.
In 2009, Google revealed that an internet search consumes about 1 kilojoule of energy. Industry experts estimate that by 2016, global data centres – a huge whole of servers and switches – were consuming almost 40% more electricity than the entire UK.
After that, this number is likely to grow as the volume of data far outweighs the increase in efficiency.
Another challenge for a data-intensive world is that even at the consumer device level, data transfer rates are beginning to exceed the capabilities of traditional interconnect technologies. For example, the ultra-high pixel density and high frame rate of the latest HDTVs make it increasingly difficult for traditional copper HDMI cables to function. Even over relatively short distances in home entertainment systems, the degree of signal degradation of such cables is very noticeable.
At the heart of the problem is that traditional electronic data systems need to charge and discharge wires in order to send data from point A to point B. Even for the tiny wires inside the CPU and RAM chips, this charge-discharge cycle requires energy and time.
As David Miller, an applied physicist and electrical engineer at Stanford University in the United States, has pointed out, most of the energy used in information processing is used for communication, not logic. Even at the gate level, the main driver of energy dissipation is the capacitance of the charge-discharge wires, which is about 200 atofara (10 -18F) per micron wire. At the data center level, the largest server farms can consume power from an entire power plant, so the efficiency of server-to-server interconnects poses serious sustainability concerns.
Many of the problems are greatly alleviated if cables and switches are configured to communicate using photons instead of electrons
The promise of photonics in computing and communication is actually much more than that, but also includes new possibilities in logic and processing, as well as certain methods of quantum computing. These applications are near- to medium-term things, but photonic interconnect technology is already available today, with advantages such as scalability, capacity, parallelism, longer link lengths, and speeds. Even consumer apps can benefit from this.
Optical HDMI cables now support high frame rates for displays with widths of 4000 or 8000 pixels (or even wider) on commercially useful cable lengths, while optical USB and Thunderbolt data transfer rates are reaching 40Gbps and beyond.
Quick Solution: PI's "Parallel Digital Gradient Search" firmware enables silicon photonic system designers to quickly adjust optical components for maximum power transfer. The optimization process takes only 1% of the time of the old-fashioned sequential technique
However, manufacturing these photonic devices is not an easy task.
Of course, the manufacturing process starts from the same silicon wafers that have been ubiquitous in microelectronics for decades, using many of the manufacturing tools found in standard semiconductor fabs to print optics and circuits onto wafers along with (or instead of) microelectronic components.
Fundamentally, photons don't follow the same principles of physics as electrons. Directing photons from one component within a packaged "chip" to another is not as simple as soldering the two together.
First of all, alignment tolerances are much tighter;Attaching wires to contact pads on a chip requires aligning components to within a few tens of microns of the correct position, while attaching optical fibers to photonic chips requires three orders of magnitude more accuracy. Fortunately, thanks to ingenious device engineering and some groundbreaking microrobotic industrial automation technologies, scientists are overcoming these challenges and bringing high-throughput photonic interconnect technology into the mainstream.
The challenges of manufacturing systems that utilize photons to transmit information are not new.
In fact, these challenges date back at least to the late 90s of the 20th century, when photonic technology was first applied on a large scale to replace satellite links in long-distance telecommunications. At that time, the delivery of light from the (then) novel laser diode into a single-mode fiber required laborious adjustment of the position of the two elements before the light of the laser could be effectively coupled into the fiber.
Part of the solution is an analog technique called gradient search, which allows the positioning system to quickly reach the optimal transmission value (at least for fibers with smooth modal profiles).
However, the positioning system itself has significant limitations. They are often fragile, have limited range of movement, and are prone to deviation from the alignment – a quality that is not a good quality for a system deployed in an industrial environment. This problem was finally solved by developing a digital version of gradient search and deploying it on industrial-grade motion hardware. The combination of the two provides a way forward for manufacturing rugged fiber optic interconnect hardware.
Today's photonic data devices are more sophisticated and complex. Thousands of photonic integrated circuits (ICs) can be fabricated on a single wafer, offering an attractive potential to process multiple optical channels simultaneously, increasing capacity and speed. Photonic integrated circuits typically integrate multi-channel and multi-wavelength structures, with inputs and outputs arranged in arrays that enable each device to process and transmit multi-channel information.
Today, moving data at the speed of light means using specialty materials. These include indium phosphide (INP), the standard for lasers and other technologies that push photons on fibers, and silicon germanium (Sige), which is widely used in high-speed mixed-signal electronics to control light. But getting all these building blocks into inexpensive silicon wafers that can be produced at scale is a huge challenge, which comes with a unique set of problems compared to cramming more transistors into a CPU.
In some ways, getting these circuits to "talk" to each other is the same challenge as telecom fiber. All chip-scale photonic inputs and outputs need to be coupled to other components, including optical fibers, fiber arrays, waveguide-based structures, bulk components such as laser diodes, lenses, and gratings, and other chips. Many of these couplings require fine alignment not only on the sensitive transverse plane (Xi referred to as the "xy" plane), but also on other degrees of freedom. Importantly, the photonic device array must be precisely positioned not only in the xy plane, but also in the z-axis direction and, in general, in other degrees of freedom (DOFs).
Another complication is that the necessary alignment can only be actively made by measuring the actual coupling of light to fiber. This is because machine vision methods typically used to determine device position during assembly cannot address spatial tolerances in the tens of nanometers, while passive alignment methods, such as precision V-grooves that glue fibers in place, often require unrealistic device replication tolerances.
In principle, digital gradient search in the 90s of the 20th century can still be used for the alignment of optical components in silicon photonics. However, in practice, due to the large number of array elements involved and the need for multi-dimensional optimization, the process of finding the global optimal alignment is not simple. For example, moving a component on the -z axis will inevitably cause it to be misaligned in the xy plane because the mechanical axis of rotation does not exactly coincide with the optical axis.
This problem is usually solved using the "loop" method: aligning the components on the xy axisMinor improvements on the Theta-Z axis;Return to the xy axis and realign;This is then repeated until the results are satisfactory. The round-robin approach can also be used to optimize devices such as short multimode waveguide structures commonly found in silicon photonics, where there are interactions between channels and between inputs and outputs.
In this case, the input is optimized first, and then the output: only now the input is no longer optimized, so the cycle is repeated until a consistent optimal value is reached. Once the optimization process is complete, the digital gradient search technology tracks the arrangement of the components, ensuring that the components remain optimized under the effects of thermal changes and stresses on the solidified glued elements.
But there's a problem: the cycle is too slow and impractical for testing and subsequently assembling thousands of devices from each wafer. To understand why, consider the early stages of wafer fabrication.
As with traditional microelectronics manufacturing,The manufacturing cost of silicon photonic chips is only a fraction of the finished product of the packaged chip**。Since the wafer yield is by no means 100%, it is economically beneficial to discard the problematic device before it enters the packaging stage, which is much more costly.
In microelectronics manufacturing, this quality control process is done using specialized tools called wafer detectors, which typically use precision instruments on a shelf that bring each chip into contact with a precise needle-like electrode, stimulate the chip, and observe its reaction.
However, to perform the same test on a photonic chip, electrical contact alone is not enough. The embedded photonic circuitry of the chip must also be optically coupled into and out of the chip, and the performance of the chip must be measured optically. Array elements present special challenges as these devices must be rotated to a precisely matched orientation and converted into a tight transverse arrangement. Traditionally, such optimizations have been done gradually, with angle adjustments staggered with lateral realignment to compensate for any mismatch between the optical axis and the axis of rotation. This can take several minutes, which is unacceptable (and uneconomical) considering that each wafer may contain thousands of silicon photonic chips.
Recently, a new digital gradient search method has emerged that enables industrial locators to perform gradient searches on multiple channels, inputs, and DOFs simultaneously. Instead of repeating small motions of -z and xy correction staggering over a matter of minutes, this technique can align xy and -z at the same time, using two gradient search processes in parallel. This Parallel Digital Gradient Search (PDGS) method can be extended to all six degrees of freedom per device, allowing multiple locators to work together on multiple inputs and outputs of a device, even if there is interaction between inputs, outputs, and channels.
With PDGS, the minutes-long serial process of cyclic (e.g.) XY and Theta-Z alignment of array devices takes only about a second, saving 99% of the time.
The parallelism of the technology also means that the time of the entire process is almost independent of the number of adjustments performed. This is significant because in the manufacturing process (as with many things), time is money. Not surprisingly, manufacturers and users of wafer probes pioneered the adoption of miniature positioning robots using PDGS, which were first implemented in 2016. Microrobots are also used in other areas of the manufacturer's assembly process, such as aligning chips and other components together to package them. Some manufacturers have even adopted specialized test tools to re-verify the health of the chip in the middle of the packaging process.
This parallel process is also used in the firmware of industrial-grade hexapod microrobots and other precision machinery used in the manufacturing of silicon photonic devices, thus changing the unfavorable economic conditions of the silicon photonics industry.
Silicon photonics is a technology that utilizes silicon-based materials to manipulate and transmit light to create low-latency computing and interconnect solutions. One of its main advantages is the ability to integrate optics and electronics into a single package;This makes it possible to create highly integrated systems: they can process both electronic and optical signals, making it possible to create more efficient and compact devices.
There are many misconceptions about how photonics can be used to improve today's designs. People understand the potential of photonics, but is it ready for use?
In general,The value of silicon photonics technology lies in its potential to enable new technologies and applications that are not possible with traditional electronic circuits, as well as to improve the efficiency and performance of existing technologies. So, what "rumors" have appeared?We will ** 11 of them and present evidence to debunk them.
1) Nice research project, but not commercially viable at the moment
In recent years, silicon photonics technology has attracted attention due to its potential to provide high-speed data transfer, increased memory bandwidth, and low power consumption. While there are still some challenges to the implementation of silicon photonics, it has shown good commercial viability in a variety of applications.
Lightelligence's HummingBird ONOC platform enables innovative interconnect topologies with silicon photonics technology to improve computing performance. Its waveguides propagate signals at the speed of light and utilize an all-pair, full-datacast network to transmit signals to each core on a 64-core domain-specific AI processor chip.
One such application is data center interconnect, which can provide high bandwidth, low latency, low power consumption, and rack-to-rack connectivity. If a repeater is not used, the high-speed copper connection is limited in length. Silicon photonics technology improves data center interconnectivity by providing lower-latency compute high-speed link (CXL) connectivity between servers, GPUs, and memory pools.
Silicon photonics technology uses a standard CMOS manufacturing process, which reduces manufacturing cost and complexity, making it a more viable commercial solution.
2) CMOS will continue to expand according to Moore's Law
Since the dawn of the semiconductor industry, the increase in computing power can be described by Moore's Law: transistor density doubles every 18 months, allowing CMOS chips to continuously increase computing power while maintaining constant energy consumption and area consumption.
As chip manufacturing processes move to 5nm and 3nm, transistor density is approaching its physical limit. Moore's Law is slowing down, and the traditional way to increase computing power on a single chip is no longer sustainable.
Silicon photonics technology will allow for continued expansion, beyond the limits of Moore's Law. Currently, digital chips are limited by the physical limits of the underlying component: the CMOS transistor.
Optical signals and devices follow different physical principles: the interaction of optical signals is usually linear and can be mapped to linear calculations. Over time, they will enable optical computing accelerators to outperform CPUs and GPUs.
3) The use case for silicon photonics is very narrow
Silicon photonics is a technology that uses silicon as a platform for the generation, manipulation, and detection of light. It is a proven technology in telecommunications and data center applications where information is transmitted over fiber optic connections.
Now, it has been applied to other computing needs where system scaling requires high bandwidth and low latency, and has the potential to revolutionize a wide range of industries.
Some of the applications include:
Data center interconnects for the high-speed transmission of large amounts of data over long distances. - High-performance computing, which connects multiple processors together for faster data transfer and reduced IO bandwidth for large memory pools. - Telecommunication networks that increase data transmission capacity and speed. - Lidar systems for autonomous vehicles and other autonomous vehicles for accurate distance measurement. - Quantum computing, which is used to control and manipulate qubits, opening the door to faster and more efficient quantum computing.
4) Photonics requires a major shift in methodology
Utilizing silicon photonics requires extending existing methods from traditional electronics to photonics. In electronics, information is transmitted and processed electronically. In photonics, information is transmitted and processed with light: fabricating photonic integrated circuits requires a different set of design rules, manufacturing processes, and test methods.
The use of photonic integrated circuits in the system makes full use of existing methods and design rules. Many photonic integrated circuits use electronic integrated circuits in a common package configuration to convert electrical signals into optical signals, which in turn convert optical signals into electrical signals. Since the optical signal is internal to the device, this application is transparent to the system designer.
5) Silicon photonics technology applications require high Xi curves
The Xi curve for manufacturing silicon photonic devices is not as steep as other emerging technologies. Silicon photonics technology involves the integration of multiple disciplines, including electronics, optics, materials science, and manufacturing technology.
The use of silicon photonic integrated circuits is similar to electronic integrated circuits in that photonic integrated circuits are usually co-packaged with electronic integrated circuits. These integrated circuits typically have electrical interfaces and behave similarly to other integrated circuits in system design.
6) Digital integrated circuits and photonic integrated circuits are difficult to integrate
Digital integrated circuits and photonic integrated circuits can be integrated, and some companies have adopted a hybrid integration approach when producing photonic integrated circuits to ensure their reliability and performance.
The hybrid integration approach is commonly used to integrate electronic and photonic integrated circuits by fabricating digital and photonic circuits separately and then bonding them together. This can be achieved through various techniques such as flip chip bonding, wire bonding, or solder bonding.
The design and manufacture of hybrid integrated circuits requires careful consideration of several factors, including thermal management, electrical and optical coupling, and packaging. Since digital circuits generate a lot of heat, which affects the performance of photonic circuits, thermal management is critical to ensure the reliability and stability of integrated circuits.
7) Artificial Intelligence Machine Xi (AI ML) workloads will continue to scale on GPUs
Artificial Intelligence Machine Xi (AI ML) workloads can be scaled on GPUs without the use of photonic technology, and are the workhorse of AI Machine Xi (AI ML) training due to their ability to handle parallel processing tasks, which is critical for training large models. Thanks to recent improvements in GPU performance, they can handle increasingly complex AI ML workloads.
However, as AI ML workloads continue to grow in complexity and scale, there is a need for faster and more efficient data transfer between GPUs and other processing units. This is where photonics can come into play, enabling high-speed, low-latency interconnects between GPUs, CPUs, and other processing units, increasing efficiency and data transfer speeds.
As AI ML models become more complex, specialized hardware accelerators such as tensor processing units (TPUs) and field-programmable gate arrays (FPGAs) will be required. These accelerators can use silicon photonics technology for efficient interconnection and data transfer through low-latency CXL connectivity.
In addition, new photonic accelerators use low-latency optical NOCs (optical networks on chips, or ONOCs) to improve the performance and throughput of AI workloads. In terms of specific features, these accelerators are up to 800 times faster than state-of-the-art GPUs.
8) 3nm digital chips perform better
It is difficult to directly compare the performance of 3nm digital chips and silicon photonic chips because they are designed for different applications and use different metrics to measure performance. The choice depends on specific application requirements and design considerations.
3-nanometer digital chips are designed to process digital information such as performing arithmetic operations, logic functions, and memory operations. They are optimized for high-speed computing and low power consumption, which is critical for applications such as mobile devices, data centers, and high-performance computing.
Silicon photonic chips are mainly used to transmit and process optical signals. They are optimized for high-speed data transmission and low power consumption, which are critical for applications such as data communication, sensing, and medical imaging.
While 3nm digital chips may outperform silicon photonic chips when it comes to digital processing tasks, silicon photonic chips have the potential to provide higher bandwidth and lower latency for long-distance data transmission. For example, photonic technology can be used to improve access to memory by digital chips, greatly reducing memory bottlenecks. In addition, silicon photonic chips can provide higher energy efficiency for certain applications where power consumption is critical.
9) Optical NOC is too small to be practical
Optical NOCs have the potential to provide high-bandwidth, low-latency communication between on-chip processing units. Usability depends on various factors such as cost, power consumption, reliability, and scalability.
The cost of implementing optical NOC requires specialized components such as lasers, modulators, and detectors, which can be expensive to manufacture and integrate. While optical communications consume less power compared to electronic communications, optical components such as lasers and modulators still consume a lot of power.
Optical NOCs offer several key advantages over digital chips. Since the connection between the digital chips uses light, the optical NOC has extremely low latency and can communicate with all nodes of the entire device at the same time. This solves the problem of the nearest neighbor of the digital integrated circuit, allowing new topologies to be considered.
Unlike digital NOCs, optical NOCs are not limited to a single grille. With wafer splicing, optical NOCs provide wafer-level density when using standards such as Universal Chip Interconnect Express (UCIE) for chip-to-chip digital interfaces.
10) CXL is suitable for copper interconnects, not for optical interconnects
CXL is a high-speed interconnect standard designed to enable communication between CPUs, GPUs, memory, and other processing units. While the original version of CXL used an electrical interconnect, there is a growing interest in utilizing optical interconnects to achieve CXLs of more than one meter.
Compared to electrical interconnects, optical interconnects offer higher bandwidth, lower latency, and lower power consumption, especially over longer distances. The use of optical interconnects in CXLs above one meter also allows for more flexible system design and deployment. With optical interconnects, processing units can be placed farther away, allowing for more flexible system configurations and reducing the need for complex cabling.
11) Other technologies are more promising
Silicon photonics technology is being used for high-speed interconnects in computing systems, including interconnects between processors, memory, and other components. Another application is the development of photonic integrated circuits that are co-packaged with electronic integrated circuits. These packages provide a more compact and efficient solution for data transfer between computing system components, reducing the need for complex cabling and improving overall system performance and power consumption.
In conclusion, silicon photonics has become one of the key building blocks of data centers in recent years. It harnesses the power of photons to connect switches, servers, and other devices quickly and efficiently over long distances.
But as bandwidth demands continue to rise, silicon photonics is bound to become even more important. As the decline of Moore's Law takes its toll on the tech industry, companies are trying to push silicon photonics deeper into the data center.
Nvidia, for example, says it is designing GF Fotonix-based high-bandwidth, low-latency, energy-efficient optical interconnects in some "leading-edge" data center systems to handle increasingly heavy AI workloads.
One of the advantages of GF Fotonix is that the monolithic architecture reduces the rate of errors in data transmission and reduces latency by a factor of 10, which in turn brings higher throughput to AI workloads.
To make its job easier, Nvidia has partnered with Ayar Labs, a startup that designs optical interconnects that can be incorporated into a variety of processors and accelerators. This interconnection comes in the form of a chipset that can be packaged into a wide range of devices, from CPUs to GPUs, providing 1,000 times more bandwidth than electrical IOs while consuming one-tenth the power of the latter.
Xanadu, a startup that relies on silicon photonics technology to enable quantum computing, has also received support from GF Fotonix.
As more and more advanced optical devices are manufactured each year, the range of silicon photonics applications will continue to expand. For example, by equipping microrobots with nanoscale eyes, this technology could optimize the position of multiple optical components in silicon photonics and other manufacturing fields, allowing precise positioning to keep pace with new chapters in computing, imaging, and many other photonics technologies.
Silicon photonics technology is attractive in various fields due to its higher bandwidth and faster communication. In addition, the surge in machine Xi (ML) and artificial intelligence (AI) applications has further strengthened the demand for silicon photonics.
The silicon photonics technology market was valued at USD 1.6 billion in 2022 and is expected to grow to USD 19.4 billion by 2032, growing at a compound annual growth rate (CAGR) of 254%。In 2022, North America held a prominent position with a 37% market share in the silicon photonics market. Experts expect the Asia-Pacific silicon photonics market to grow at a CAGR of more than 28% from 2023 to 2032.
In conclusion, silicon photonics is a beacon of innovation in computing and data communications, providing a cost-effective and energy-efficient solution for faster communication. With the development of emerging trends in silicon photonics, the field is expected to make unprecedented progress in the coming years.