If you think today’s era is on the cusp of a technological revolution, think back to the mid-1980s, when silicon chips used micron-scale transistors and fiber-optic systems sent data around the world at trillions of bits per second .
Combining the potential of silicon digital logic, optoelectronics, and fiber-optic communication technologies, anything seems possible.
As envisioned by engineers at the time, these technologies would continue to evolve and converge until photonics combined with, and eventually replaced, electronics. Photonic technology not only enables cross-border data transfer, but also transfers data between data centers and even between computers. Engineers believe that data can be transferred between chips via optical fibers, and have even envisioned photonic chips: Many hope that one day, ultra-fast logic chips will use photons instead of electrons.
However, such an idea has not been realized. Companies and governments have invested billions of dollars in research into new optical devices and systems that use optical fibers to connect racks of computer servers inside data centers. Granted, many modern data center racks are connected using these optical devices. Photonic technology, however, ends there. Inside the rack, individual server boards are still interconnected using cheap copper cables and high-speed electronics. Of course, on the circuit board, it's all metal wires that connect the processor.
Attempts to push photonics technology into the servers themselves, with optical fibers directly connecting processors, hit the rocks and failed for economic reasons. According to the market research firm Light Counting, the market size of Ethernet optical transceiver equipment has reached an average of 4 billion US dollars a year, and by 2020 this market will expand to nearly 4.5 billion US dollars and 50 million sets of devices, which is undeniable . But to this day, photonics has not solved the problem of the last few meters between the computer rack and the processor chip in the data center.
Still, the huge potential of photonic technology means there is still hope. While the technical challenges remain enormous, new ideas in data center design offer a plausible path to the photonics revolution in the era of big data.
In today's digital world, whether to surf the Internet, watch Internet TV, or do anything else, the data streams that people use go through optoelectronic transceivers. The function of the photoelectric transceiver is to realize the mutual conversion of photoelectric signals. Inside the data centers of major cloud service providers and social media companies, these transceivers sit at the end of the fiber used to transmit data between data centers. Transceivers connect to switching equipment at the top of each server rack, converting optical signals to electrical signals and transmitting them to the server group in that rack. Transceivers also convert data from servers into optical signals and transmit them to other racks, or via network switches to upload to the Internet.
Each optoelectronic transceiver consists of three main components: a transmitting device containing one or more light modulators; a receiving device containing one or more photodiodes; and a complementary metal-oxide-semiconductor (CMOS) logic chip that encodes and decodes data . Since ordinary silicon is not suitable for light, the photons originate from a laser that is separated from the silicon chip (although the laser can be packaged with the silicon chip). In this way, instead of switching the laser on and off to represent the bits, the laser remains on and the electronic bits are encoded into the laser by the optical modulator.
The optical modulator is the core of the transmitting device, and there are many types. There is a particularly small type of modulator known as a Mach-Zehnder modulator. The modulator transmits the laser light through a narrow silicon waveguide. The waveguide splits in two and then merges back into one after a few millimeters. In general, such shunting and merging will not affect the light output because the two branches of the waveguide are of equal length. When the branches are combined into one, the phases of the two light waves are still the same. However, applying a voltage to a branch changes its index of refraction, effectively reducing or increasing the speed at which light waves travel. This way, when the two light waves meet again, they destructively interfere with each other, suppressing the signal. Therefore, changing the voltage of a certain branch is actually using the electrical signal to modulate the optical signal.
The construction of the receiver is much simpler, including a photodiode and some support circuits. The optical signal is transmitted via an optical fiber to a germanium or silicon germanium diode in the receiving device, which generates a voltage with each light pulse.
Both the sending device and the receiving device are provided with circuits, through which the signal amplification, data packet processing, error correction, buffering and other tasks are performed, so that the optical fiber conforms to the standard of Gigabit Ethernet. Whether these devices are on the same chip or the same package as the optical components varies by vendor. But most electronic logic and optical components are separate.
With the increasing use of optical components on silicon integrated circuits, you might think that direct integration of optical technology with processor chips is an inevitable trend. Indeed, it did seem that way for a while.
But in fact, people completely downplay or even ignore the fact that there is a growing gap between the speed at which the smallest feature size of electronic logic chips shrinks and the ability of photonic technology to keep pace with it. The feature size of today's transistors is only a few nanometers. Using 7-nanometer CMOS technology, more than 100 general-purpose logic transistors can be integrated in an area of ​​1 square micrometer on the chip, and this does not include the maze-like complex copper connection lines on the transistors. On each chip, in addition to billions of transistors, a dozen layers of metal wiring are required to connect these transistors to form registers, amplifiers, arithmetic logic units, and the complex units that make up the processor core and other vital circuits.
The problem is that a standard optical component, such as a modulator, must be no smaller than the wavelength of the transmitted light, which limits the size to about 1 micron wide. Moore's Law can't solve this problem, nor can more and more advanced lithography. This is simply because electrons are very "thin" with wavelengths of just a few nanometers, while photons are relatively "fat".
That being the case, can't chipmakers simply integrate modulators and reduce the number of transistors? After all, there are billions of transistors on a chip today. The answer is: no. Silicon electronic chips carry a huge amount of system functionality per square micrometer, and even replacing a small number of transistors with lower-functioning components, such as optical components, can be very expensive.
The specific calculation method is as follows. If there are an average of 100 transistors on a chip per square micrometer, and an optical modulator occupies an area of ​​10 micrometers by 10 micrometers, it will replace a circuit containing 10,000 transistors! Remember, a standard light modulator is just a simple switching device that turns the light beam on or off, and each transistor acts as a switch to switch current. Thus, a rough estimate is that the opportunity cost of this fundamental switching function is 10,000:1 for photonic components compared to electronic components, since the system designer has at least 10,000 electrons in the latter versus one light modulator in the former. switch can be selected. Even if the direct integration of the modulator with the processor could improve performance and efficiency, no chipmaker would accept such a high cost.
There are other obstacles to replacing electronic components on a chip with photonic components, such as photonic components that do not provide some of the key functions that chips require (such as storage). The conclusion is that photonic components are incompatible with basic computer chip functionality. Even if this were not the case, it would be pointless to replace electrons that perform the same function with photons.
That's not to say photonics can't get any closer to processors, memory, and other vital chips. At present, top-of-rack (TOR) switches equipped with optical transceiver modules are attracting the attention of the data center optical interconnection technology market. The top-of-rack switches are installed on the top of a 2-meter-high rack, and there are server chips, memory and other equipment resources in the rack. . Fibers are interconnected with other TOR switches through a separate switching layer. These switches are then connected to another set of data center gateway switches to connect to the Internet.
A typical TOR panel with embedded transceiver modules provides insight into the amount of data it transmits. Each TOR switch is connected to a transceiver module, and each transceiver module is connected to two optical fibers (one for sending and one for receiving). 32 modules can be embedded in each 45mm high TOR panel, and each module can transmit data at speeds of up to 40 gigabits per second in each direction. In this way, the data transfer speed between the two racks can reach 2.56 terabits per second.
But the data flow inside the rack and inside the server still uses copper wire. Unfortunately, this has become an obstacle to building faster, more energy-efficient systems. Using photonics solutions to solve the last meter (or two meters) of server or processor connections may be the best opportunity to develop a large-scale optical device market. But before that, there are some challenges in terms of price and performance that need to be overcome.
Fiber to the processor solutions are not new. There have been many attempts in the past, and many lessons have been learned in terms of cost, reliability, power, and bandwidth density. For example, about 15 years ago, I was involved in designing and building an experimental broadband transceiver. The experiment wanted to connect 12-fiber-wide parallel fiber ribbons into the processor. Each fiber transmits a digital signal, which is generated by four vertical-cavity surface-emitting lasers (VCSELs). The laser is a diode that emits laser light from the surface of the chip, producing a much denser laser than edge-emitting lasers. Four vertical-cavity surface-emitting lasers encode information by switching the optical output, and they transmit at different wavelengths in the same fiber, this coarse wavelength division multiplexing technology can quadruple the capacity of the fiber. Therefore, the data transmission speed of each VCSEL can reach 25 gigabits per second, and the total system bandwidth can reach 1.2 terabits per second. According to the current industry standard, 12 fibers are arranged side-by-side with a gap width of 0.25 mm between adjacent fibers, so the bandwidth density is about 0.4 Tbit/sec/mm. In other words, a 1mm-wide optical fiber can process the information stored by the US Library of Congress Network Archives team for a month in 100 seconds.
The data speeds required for fiber access processor applications today are even higher than this, but this is a good start. So why wasn't this technology used at the time? Part of the reason is that the manufactured systems are both unreliable and infeasible. At that time, it was difficult to manufacture the 48 vertical cavity surface lasers required for the transceiver, and it was difficult to guarantee that the transceiver would not fail during its service life. In fact, an important lesson is that designing a single laser with multiple modulators is more reliable than 48 lasers.
Today, the performance of vertical cavity surface emitting lasers has improved, and transceivers based on this technology can effectively provide short-range solutions suitable for data centers. In addition, fiber ribbons can be replaced with multicore fibers. Multi-core fibers can split the same amount of data into multiple fiber cores embedded in the main fiber. Another recent development is the release of more sophisticated digital transmission standards, such as PAM4, which encodes information with four levels of light intensities instead of two, increasing the speed of data transmission. There are also scientific projects (such as the Shine project at MIT) that are working to increase the bandwidth density of the fiber access processor demonstration system to 17 times the value we achieved 15 years ago.
But these advances combined are still not enough to make the application of optical technology in processors further. However, I think that this step will be realized sooner or later with the emerging data center system architecture transformation.
Today, blade servers in data center racks have processors, memory, and storage. In fact, it does not have to be the case. The memory is not placed together with the server chip, but is placed on a separate rack, or even on a different rack. The Rack Architecture (RSA) is believed to allow more efficient use of computer resources, simplifying the task of hardware management and replacement, especially for large social media companies like Facebook, where the amount of computation required for special applications and memory will keep growing over time.
Why is this architecture enabling a breakthrough in photonic technology? Because a new generation of efficient, inexpensive, and terabit-per-second optical switch technology can achieve just that dynamic resource allocation and reconfigurability.
The main obstacle to this approach to data center redevelopment is the price of the device and the cost of production. Silicon photonics already has a cost advantage in leveraging off-the-shelf chip production lines and taking advantage of silicon's vast infrastructure and reliability. But the combination of silicon and light is not perfect: silicon emits light inefficiently, and silicon has severe optical losses. A standard silicon photonics transceiver produces at least 10 dB (90 percent) of optical loss by measuring the optical input and output. But this inefficiency doesn't affect short-range connections between TOR switches, because the potential low-cost advantages of silicon outweigh the problems, at least for now.
A major cost of silicon photonics modules comes from the humble but important optical connectors. The optical connection includes not only the connection between the optical fiber and the pickup device chip, but also the connection between the optical fibers. Hundreds of millions of extremely high-precision fiber optic connectors must be produced each year to meet demand. How high is the specific accuracy? The single-mode silica glass fibers used in optical connectors are 125 microns in diameter, slightly larger than the diameter of a human hair. The precision that this single-mode fiber in the connector must achieve is 100 nanometers, which is only 1/1000 of the diameter of a hair, otherwise the signal will be greatly weakened. The production method of optical connectors between optical fibers and between optical fibers and transceivers still needs further innovation to meet customer demands for precision and low cost. However, there are few production technologies that can meet this demand for good quality and low price.
One of the ways to reduce costs is, of course, to reduce the price of chips in optical modules. There are many ways to produce chips, but a technique called wafer-scale integration helps keep costs down. Silicon wafer integration technology is to make photons on one silicon wafer and electrons on another silicon wafer, and then glue the two silicon wafers together. The paired wafers are then diced into chips and fabricated into nearly complete modules. (Lasers made from non-silicon semiconductors are kept separate.) This approach allows assembly to run in parallel with production, reducing costs.
Another factor that reduces costs is of course the scale of production. Assume that the total size of the optical Gigabit Ethernet market is 50 million transceivers per year, and the area of ​​each optical transceiver chip is 25 square millimeters; then assume that the factory uses silicon wafers with a diameter of 200 mm for production, and the output rate To reach 100%, then the number of silicon wafers required is 42,000.
That sounds like a lot, but it's less than two weeks of production in a standard factory. In fact, even if a transceiver manufacturer has a quarter of the market, it can only maintain production for a few days. To reduce costs, you need to increase production. The only way to do this is to apply optical technology to the processors inside the servers beneath the top-of-rack switches.
For silicon photonics to seize opportunities in existing all-electronic systems, it must have strong technical and commercial advantages. Such a device must solve a major challenge and significantly improve the overall system. It also had to be small, energy efficient, extremely reliable, and had to transfer data at breakneck speeds.
There are currently no solutions to meet these requirements, so electronics will continue to evolve without having to be tightly coupled with photons. Without major breakthroughs, "fat" photons still have no place in functional systems dominated by "thin" electrons. However, if optical components can be reliably produced in large quantities and at low cost, decades of vision for fiber access processor technology and related architectures can become a reality.
We have made a lot of progress in the past 15 years. We have a better understanding of photonics and a better understanding of where it is available and where it is not available in the data center. In addition, a sustainable, multi-billion-dollar-a-year commercial market for optical devices has emerged. Optical connectivity has become an important part of the global information infrastructure. But at present, it is not feasible to apply a large number of optical devices in the core part of the existing electronic system.
So will this always be the case in the future? I don't think so.
Solar Home Lighting System Kit
SHENZHEN CHONDEKUAI TECHNOLOGY CO.LTD , http://www.siheyidz.com