Silicon's rapid rise in Radio Frequency or RF applications is a lesson in the power of Moore's Law. Over twenty years ago, all the parts of an RF front end module were made separately and integrated in a box. Silicon changed all that. And helped make possible today’s slick and thin cellphones. In this video, the unexpected rise and quizzical future of RF-CMOS. ## The RF Front End Let us begin with RF chips. These are microelectronics that operate within a frequency range of about 100 megahertz to 100 gigahertz. Or more. The chip's front-end module is made up of a set of components that sends and receives signals. Let us briefly walk through it. At the start, we have an analog signal of our information, called the baseband. Most information signals are low frequency. Our human voice for instance is about 80 to 200 hertz. Trying to send such a signal as is however would need a large antenna. Since a low frequency means a long wavelength, and the rule of thumb is that a good antenna should be half the wavelength. An antenna for a 200 hertz signal would be hilariously, impractically big. So we mix this low-frequency baseband signal with a higher frequency signal called a carrier frequency. This signal would be generated by a local oscillator. Oscillator, not ocelot. The mixing would be done by a device called a "mixer". With this mixing, the baseband is upconverted into a passband signal. Its power is then amplified by the Power Amplifier before it is sent to the antenna for transmission. The process of receiving the signal is the largely the same but in the reverse direction. After the antenna picks up the signal, a Low Noise Amplifier component amplifies it - hopefully without adding too much noise. Mixers then downconvert the signal back from the passband to the baseband. Sometimes, another set of mixers are added to go to an intermediate frequency between the pass and basebands. This whole system is called a superheterodyne transceiver. "Heterodyne" meaning the mixing of different signals. “Super” because the signal frequencies involved are very high. And "transceiver", because the transmit and receive functions are combined into a single device. ## Good RFs These various components are made from several devices like transistors. What makes a transistor suitable for an RF semiconductor? First, it has to have a high cutoff frequency - which refers to the highest frequency signal that a transistor can still effectively amplify. An old rule of thumb is that the cutoff frequency should be 10 times higher than the maximum operating frequency. So a transistor with a 200 gigahertz cutoff frequency can be used for ICs operating in frequencies up to 20 gigahertz. Ratios of 3-5 have also been proposed before. The next concern is noise. Noise interferes with our ability to read the analog signal. When you run a signal through an RF transistor, the transistor should add as little noise to that signal as possible. Another important metric is power. The signal loses power as it travels out from the user. So we need to deliver as much output power to the signal as possible, so to allow the receiver to receive a signal with little error. ## BJTs Digital Logic is the largest market in semiconductors, 70% of the whole industry, and it is all dominated by silicon Metal Oxide Semiconductor Field Effect Transistors or MOSFETs. But this dominance is a bit of an anomaly. In the early years, research was conducted on all types of materials and devices - particularly compound alloys known as III-V semiconductors. This was especially the case in RF, because silicon MOSFETs could not switch fast enough to operate at the gigahertz frequency ranges RF devices require. Soon after the invention of the first transistor, work was done to improve their operating frequencies. In 1958, a Germanium-based bipolar junction transistor or BJT for the first time ever managed to amplify a gigahertz signal. An ordinary BJT has three sections of doped semiconductor material - an emitter, base, and collector. When a forward voltage is applied, electrons are injected across the base from the emitter to the collector. ## MESFETs Then in 1965, an RF breakthrough by the legendary Carver Mead at CalTech. Mead coined "Moore's Law" term, first proposed the independent foundry concept, and co-invented VLSI design principles alongside the late, great Lynn Conway. To add to that amazing legacy, in 1965 Mead fabricated the first Gallium-Arsenide metal-semiconductor field effect transistor or MESFET. Like the far more well-known MOSFET, the MESFET has a source and drain too. But unlike the MOSFET, the flow of charge carriers between them is directly controlled by a metal gate called a Schottky metal gate. In other words, the gate sits right on top of the channel. This differs from the MOSFET, which has a layer of oxide - the gate oxide - sitting in between the gate and the channel. The MESFET's key advantage over the MOSFET is that charge carriers can very quickly speed through the channel. In part because it is made from Gallium-Arsenide, a III-V semiconductor material with far superior electron mobility than silicon. 8,500 square centimeters per volt-second as compared to silicon's 1,400 square centimeters per volt-second. By 1971, MESFET gate lengths had reached about a micrometer, allowing for a cutoff frequency of 10 gigahertz. A 10 gigahertz cutoff frequency means a 1-gigahertz operating frequency, good enough to amplify microwave signals. Two years later, Hewlett-Packard released the first high-speed Gallium-Arsenide logic gate. Shortly afterwards Plessey, the British semiconductor-maker, released the first Gallium-Arsenide Low Noise Amplifier as a monolithic integrated circuit. ## Silicon CMOS Let us now talk CMOS. CMOS stands for Complementary Metal Oxide Semiconductors. And it has that name because it is made up of two different types of silicon MOSFETs side by side. One MOSFET - called a P-channel MOS or PMOS - is doped to transport electron-holes from the source to the drain. The other is a N-channel Metal Oxide Semiconductor or NMOS which transports electrons. The CMOS arrangement was invented in 1963 by Frank Wanlass and Tom Sah of Fairchild Semiconductor, based on groundbreaking MOSFET work done by RCA. They noticed that their CMOS arrangement drew close to zero power when in standby, just whatever leakage power there might be. In those days, most American semiconductor makers produced their circuits from PMOS or NMOS, but not both. For example, the iconic Intel 4004 was made with PMOS. Japanese calculator-makers in the early 1970s on the other hand selected CMOS to make the chips for their calculators. Mostly to take advantage of its power-saving nature. But as those power savings became more significant in later years, the various American semiconductor-makers switched to CMOS too. In 1985, Intel used CMOS for their breakthrough Intel 386 CPU. And ever since then, CMOS has been the workhorse of what we call digital logic systems. Intel and other companies have worked to shrink the transistors so to stuff and integrate more of them into a CPU. More transistors, better CPU. Nice. ## MESFETs and BJTs in RF Very quickly, silicon MOSFETs conquered areas once held by Germanium - including the BJT. But RF remained a stronghold. In the early 1980s, the first civil applications for RF devices hit the market - things like satellite TV, radio telescopes and early cellular devices. These electronics were made from either silicon/silicon-germanium BJTs or Gallium Arsenide MESFETs. Silicon BJTs were used for RF devices at operating frequencies below 4 gigahertz, thanks in part to their ability to output good power. And Gallium Arsenide MESFETs were given the nod when frequencies between 4 and 20 gigahertz were needed. MESFETs also outperformed Silicon BJTs in metrics like noise generation. Some foundries like STMicroelectronics offer Bipolar CMOS or BiCMOS, made with silicon-germanium. This integrates together the Bipolar Junction and CMOS transistors. The idea is to offer the best of both worlds, and was an early enabling technology for RF ICs. As the cellular industry started to take off with 2G wireless in the early and mid-1990s, companies had no time to research and develop better solutions than what was already at hand. So typical RF front-ends in that era featured various discrete components made from silicon CMOS, silicon BJTs, or Gallium-Arsenide MESFETs. This left you with rather bulky looking devices that did not garner much attention from the consumer market. ## The Rise of Silicon For decades, it seemed like silicon MOSFETs were unsuitable for RF applications. People believed that the MOSFET gate length could never be reduced below 100 nanometers, and the gate oxide, 3 nanometers. This correlated to a cutoff frequency of about 200 megahertz, making CMOS unsuitable for RF. And that was the standard trash talk from university professors and analysts deep into the 1980s. One unnamed MIT professor said: > RF is a solved problem. And using an inferior technology like CMOS to solve it yet again is stupid-squared This professor was not entirely wrong for the time. In the 1980s, the peak NMOS transistor - made with 2 micron technology - had a cutoff frequency of just about 1 gigahertz. But Moore's Law busted through all that. Transistors got smaller, letting them switch faster and giving them increasingly faster cutoff frequencies. By the early 1990s, IBM was even speculating at the time about 100 gigahertz chips, though they were talking about CPUs rather than RF. As the technical opportunity for an RF-CMOS industry started to emerge, university researchers around the world began exploring the various circuit blocks needed for an RF CMOS mixed signal integrated circuit. The dream was that radios might go the same way as switching networks in the 1980s and wireline modems did in the 1990s. Meaning being able to produce together all the analog RF front end components plus the Digital Signal Processor. Thus giving you mixed-signal ICs that are reliable, physically small, and can be cheaply produced at scale. ## The Suspended Inductor Because it began as a digital logic technology, CMOS was not particularly suited for RF. Even after the transistors shrank enough to operate at the necessary frequencies, more had to be done. Unlike Digital logic systems, which are mostly made up of transistors, RF electronics also need passive devices like inductors, capacitors, varactors, and resistors. A notable obstacle was how to make these devices using CMOS process nodes. Take the spiral inductor. These cool-looking passive devices look like coiled snakes and are made from metal. They play various important roles - including efficient signal transfer from components like the Low Noise Amplifier and the antenna. In the 60s and 70s, people believed that it was nigh impossible to fabricate these inductors on a silicon integrated circuit. To get a higher inductance value, which is necessary for these inductors to work as desired, you need to make bigger, denser spirals. However, larger spirals interacted with the silicon substrate, causing quality problems I won't get into here. That should have been that. But then in the early 1990s ever-creative CMOS researchers used a selective acid etch to literally suspend the spiral inductor above the silicon, the suspended inductor. The suspended inductor enabled what we now consider the first RF CMOS Integrated Circuit, an amplifier, created in 1993. One of the authors of that paper was Asad Abidi, a Pakistani-American UCLA professor who went on to play a major role in the commercialization of RF-CMOS. Similar things would be done for other passive items like mixers and oscillators. Look at an RF-CMOS chip nowadays and these cool looking things will jump out at you. ## Too Noisy When CMOS hit the right cutoff frequencies, critics stopped saying that CMOS was too slow and instead claimed it was too noisy. And MOSFETs do struggle in terms of noise and linearity. Can these high frequency signals pass through the MOSFET without also picking up a lot of noise? Critical to cracking this nut was the realization of "gate noise" models. First proposed in the 1960s by Aldert van der Ziel but forgotten thereafter, the model was revitalized and used to help design modern noise models that let us design a workable CMOS Low Noise Amplifier. RF-CMOS components still underperform compound semiconductor competitors. But again, the cost benefits of silicon are so much more favorable - maybe 30-40% cheaper, even when counting the variations needed for RF purposes - that you have to use it. ## RF-CMOS The rise of RF-CMOS was so significant that it ended a great deal of research being done on items like silicon BJTs and Gallium-Arsenide MESFETs. In 2001, a fabless semiconductor company called Silicon Labs released the first RF-CMOS GSM and GPRS transceiver. The company had already hit a home run in 1998 with a CMOS DAA design, a complicated chip that helps connect a modem to the wider telephone network. The CMOS version not only cost half as much as its predecessors, it took just a fifth of the board space thanks to cutting out a bunch of discrete components. The CMOS GSM transceiver does similar. It brings all the front-end components from the Low Noise Amplifier to the Mixers to a weak but usable Power Amplifier onto a single chip. 80% fewer components and 50% less space. After the release in mid-2001, Silicon Labs' stock tripled as the company shipped millions of units to handset makers like Samsung. This kicked off a massive RF-CMOS boom as companies like Infineon, Philips, and Lucent released their own single chip solutions integrating various bits. By 2002 the whole front end was done in silicon CMOS. ## Today and 5G Today MESFETs might still be used for microwave items handling very high wavelengths. But a wide variety of RF items from RFID tags to wireless LAN to GPS receivers to Wi-Fi and Bluetooth chipsets are made using CMOS silicon. And thanks to Moore's Law, we can make those RF-CMOS chips yet smaller, lighter, cheaper and in some cases faster over the years. Eventually however, RF-CMOS started hitting the same limitations that the digital logic CMOS industry hit during the early 2000s. The transistors got so small, the oxide so thin, and channel so short that it created serious power leakage issues. The introduction of FinFETs have somewhat alleviated these issues, but not completely. Regardless, RF-CMOS is still a cellular mainstay. For example, the 5G mmWave bands are between 24 and 48 gigahertz. Sample CMOS and Silicon-Germanium transceivers have existed for those bands since the early 2000s. We even have fully integrated RF-CMOS systems for automotive radars operating at the 77 gigahertz spectrum. ## Conclusion The real question is 6G. There is talk there of going beyond mmWave into the 100 and 300 gigahertz ranges, the so-called sub-terahertz bands. Can RF-CMOS technologies handle these bands? Likely not. 100 gigahertz requires a transistor with like a 500-gigahertz to 1-terahertz cutoff frequency, and the RF-CMOS record is 485-gigahertz set all the way back in 2007. Moreover, power remains a concern. At these frequencies, path loss and attenuation will be high, meaning the signal loses power faster as it travels out from the sender. To overcome that, a sub-terahertz power amplifier must pump a lot more power into the signal. And if we are talking about a handset - where power is limited and we lack active cooling systems - RF-CMOS might not be efficient enough for that. If that is indeed the case, then perhaps the future of RF-CMOS will involve going back to the past. Meaning, the use of heterogenous integration to connect RF-CMOS transceivers with power amplifiers made with III-V semiconductors. In other words, RF chiplets. Now, there's no guarantee that 6G sub-terahertz comes to pass. After all, 5G mmWave hasn't exactly set the world on fire. But if 6G does indeed open up these new sub-terahertz bands, then that might also close the book on 20 years of RF-CMOS scaling.