Is it time to say goodbye to the quartz oscillator?
"We're trying to replace quartz," claims Venkat Bahl, vice president of marketing at resonator company Discera. "Internally, we call ourselves the quartz killers." The company aims to replace the quartz crystals used in practically all electronic systems today with oscillators built from silicon using micro-electromechanical systems (MEMS) techniques. And it is big business. Bahl says his company's estimates suggest that the market for timing devices is worth between $3.5bn and $4bn. "If you slice that down to the oscillator market it is worth around $1bn, and the crystal oscillator part of that is worth maybe $500m." What's wrong with quartz? It's a by-word for timing accuracy - try finding a watch that does not have the word 'quartz' stencilled onto it somewhere. And it did not take long for the delicate mechanical mechanisms of yesterday's timepieces to be replaced with a quartz electronic heartbeat.
So how come the mechanical option is coming back in, even if it is at a vastly different scale? "Why are we going to be successful? What we love about this business is that it is a low-margin business. Quartz [manufacturers] have wrung out a lot of the efficiencies available, but they're still under constant price pressure. What we have is a MEMS CMOS device with a very different cost base. Our devices are between one-thirtieth and one-fiftieth the size of quartz devices, so our cost base is dramatically smaller. Our goal is to achieve a 50 per cent gross margin."
Electronic systems have many different timing needs across a range of frequencies. A GPS receiver whose frequency reference is accurate to 100 parts per million (ppm) will never lock on to a satellite - the industry now specifies 0.5ppm for these parts - but 100ppm accuracy would be fine for a PC system clock. RF systems designers care about phase noise in their frequency references, as they are trying to pick out the in-phase and quadrature components of a signal. High-speed wired communications systems need low levels of clock jitter to sample the increasingly indistinct rise and fall of their data signals. Some designers worry about thermal compensation, while others care more about ageing effects. Everyone worries about cost.
The problem for quartz crystals is that their performance is limited by the physics of a resonating block of material. Quartz crystals can run at tens of megahertz comfortably and can offer outputs in the 150MHz range by using a third overtone of the fundamental frequency, but this may be unstable. Such crystals are usually coupled with phase-locked loops (PLLs) to produce quartz oscillator circuits (rather than simple resonators), but this can increase phase noise and jitter. And there are issues with power consumption - when almost everything else in a circuit has been put to sleep to conserve battery life, the power used by the clock source becomes a major part of system consumption.
There are also manufacturing issues with making the small, high-frequency sources necessary for today's multi-radio mobile phones. Smaller crystals are more difficultto handle, while higher-frequency crystals have to be thinner, making them more delicate and so undermining manufacturing yield. The quartz industry knows competition is coming and has been developing ways to address some of the shortcomings of its technology. Stefan Hartmann, senior manager of the quartz device division of Epson Europe Electronics, says the company still thinks quartz is the best material for clock sources because of its low sensitivity to changing temperature. "We feel that temperature stability is an issue," he says. "The frequency of a CPU clock is not critical, so you don't need extremely good temperature compensation. But there are definitely other markets where it matters, such as GPS, GSM and UMTS as well as the Bluetooth, Zigbee and RFID markets. With wired communications we need very good accuracy to achieve very high bandwidths. For example, SONET repeaters have oven-controlled crystal oscillators in them."
The problem for the quartz industry is that it is being asked to produce high frequency parts in very small packages, especially for mobile phones. "Once I go to 2mm x 1.6mm things get extremely critical if I use a mechanical process," says Hartmann. Epson's response is to start using photolithography, rather than a mechanical process, to shape its crystals. "A normal crystal is mechanically treated, cut and shaped to achieve the mechanical resonance necessary. In our process of 'quartz plus MEMS' we use purely photolithographic processes to shape the crystal to increase the frequency range we can cover. So the production yield with our process is better than that the mechanical process. "Our megahertz frequency sources are now packed in a 2mm x 1.6mm package. That's only possible if you have special processes, because with mechan--ical processes the crystal will become very fragile," says Hartrmann.
"With this process we also have the possibility to make very fancy structures, for example to etch holes into certain areas of the device to change the characteristics. "We can make very thin chips that resonate at higher frequencies, but with thick edges to make them easier to assemble. The photolithography process also has tighter tolerances than the mechanical processes, which helps improve yield, quality and product cost." MEMS, in principle, offers the opportunity to use photolithography entirely and dispense with the quartz. As most resonators need some sort of signal conditioning and compensation, there's also an opportunity to use the CMOS process to integrate the two-in-one chip.
Unfortunately, it's not that straightforward. MEMS processes diverge from standard CMOS, which is a good way to start racking up fab costs and narrowing your sourcing options. MEMS resonators are so small that unless they are very well encapsulated, contamination can affect their performance. According to Discera, the typical MEMS resonator mass is 10-13kg, so a monolayer of water atoms will shift its frequency by 100ppm. Contamination can also affect ageing performance, as well as cutting the resonator's Q factor, which is what gives it some insensitivity against process, packaging, and in-system mechanical effects. Some manufacturers do their encapsulation by bonding a wafer to the face of the finished MEMS device wafer, which can reduce the area advantages of MEMS approaches by demand---ing a large area on to which the second wafer can be aligned. According to Bahl at Discera, though, there are advantages to the MEMS approach. It makes it easy to design and build a range of different resonator shapes to achieve various properties, frequencies and Q factors.
"We haven't come across any solution which is a one size fits all. The requirements are so diverse," he says. "We use a 'free-free clamp-clamp' beam, anchored at four points to the substrate with an electrode below it to bias it with an electromagnetic field to make it vibrate." Discera's MEMS approach is to use standard CMOS with an under-etching step to release the resonator from the die, and wafer level packaging to seal the devices afterwards. "We were looking for a process with economic advantages and a sustainable future. Our MEMS-on-top approach is very simple and easy to access with very few manufacturing processes so it's very portable," he says. And the company claims it is mastering the encapsulation issue. "Over 24 months we have managed to increase the number of die produced per wafer tenfold, because of reducing the margins used to support the wafer level packaging."
The MEMS-on-top approach means that it might be possible to integrate the resonator with conditioning circuitry to produce single-chip oscillators, rather than simple resonators. But there is always a trade-off with this approach. "Single chip integration is very exciting," says Bahl. "But we're really looking for low cost, so if a two-chip solution is cheaper that's the way I'm going to go. We do have a capability of integrating on a single die but it takes extra processing steps and causes a yield hit. The area ratio between the conditioning circuitry and the resonator could be 40 to 1." Bahl believes the company's technology gives it a number of ways it can grow beyond the pin-compatible single-frequency output products it is currently selling. "We've got a lot of areas to grow, for example into high performance, multiple outputs, multiple resonators, smaller devices and other differentiators," he says. "Given our resonator we can opt to get lower power than other approaches, which would get us into the low-end of the networking market, or we could go to higher performance and sub-picosecond jitter with Q factors in the hundreds of thousands. We've already demonstrated resonators at gigahertz frequencies."
Pushing the limits
Silicon Clocks is also using a MEMS-on-top approach, using a low temperature silicon-germanium deposition step to stack its resonators on top of CMOS circuitry. Andrew McCraith, co-founder of the company, argues that there are technical limitations with a stand-alone MEMS structure in terms of the phase noise and stability that it can achieve. "As a single chip you get lower power, increased reliability, lower size and multiple references," he said. McCraith says there will be three phases to the uptake of MEMS timing devices. In the first, MEMS resonators will replace quartz crystals to offer higher frequencies and smaller packages. The second phase will be integrated timing, which will use a single die to produce a conditioned output at lower power and with increased stability. In the third phase, multiple MEMS devices will enable dynamic clocking, some redundancy and the ability to take on non-timing functions to help eliminate more PLLs. But it will take time to get to this. "The first MEMS resonators started getting some revenue, but not the wild success the companies hoped for. Just doing a MEMS resonator is not an order of magnitude change so many applications won't find it compelling enough," says McCraith. "The MEMS resonator was never going to take off. We can all make them but the performance is atrocious. The MEMS oscillator is where you get the worthwhile innovation."
In good time
McCraith says that integrating the conditioning circuitry that turns a resonator into an oscillator delivers power and performance advantages, because the circuit parasitics are reduced. Silicon Clocks has started out with a wafer-bonding encapsulation technique, to reduce the number of variations it made from a standard process. But McCraith says the company is also working on "a more novel approach that goes to size and cost". He sees MEMS clocks being used at higher frequencies, where the reliability of quartz is an issue, and to consolidate inventory, for example by producing a multiple clock output chip that could be used in a wireless handset for global markets. "We break down markets by function," McCraith says. "At the lowest level you need a crystal with the lowest cost and quality, for example a 32KHz crystal with ±25 or 100ppm accuracy. The next step up is an oscillator chip, for example a temperature controlled crystal oscillator in a cellular or GPS application, which needs tight temperature stability of ±1 to 2ppm. These are usually in the 13 to 50MHz range, and top out at 125MHz. The next step is clocks, especially for spread spectrum clock distribution with low electromagnetic interference in PCs and printers. "We're focusing on MEMS clock chips with no off-chip crystal," he added. "Many people don't understand why clock chips and crystals can co-exist.
If you want one or two frequencies, use a crystal. If you need more, use a clock chip. In MEMS resonators it is the same - a single-output MEMS resonator has lower margins and functionality." The third major player in the MEMS timing market is SiTime, which defines, releases and encapsulates its resonators during a CMOS manufacturing process on an SOI wafer. The buried-oxide layer allows the company to encapsulate the resonators by selectively etching the oxide out from under them through voids in a protective poly layer, which is later closed by growing more poly on top. The result is a completely encapsulated resonator without the use of wafer bonding. According to John McDonald, vice president of sales and marketing for SiTime, the shift from quartz to MEMS resonators "is like the change from a vacuum tube to a transistor". He claims the company's resonators are 50 to 100 times smaller than quartz resonators, and can be delivered in standard packages because of their in-die encapsulation technique.
At the moment the resonator is built by Jazz Semiconductor in Newport Beach, while a companion chip, including drive circuits, non-volatile memory, a phase locked loop, temperature sensor and compensation, is produced in a 0.18µm CMOS process by TSMC. McDonald says that integration needs to be carefully considered: "The problem is adding four mask layers to get 500 integrated devices off an 8in wafer." He says the company has been trying to persuade chip companies to integrate its resonator in their packages, and has already sold trial lots of dice to three chip companies. Jazz Semiconductor has also demonstrated the integration of CMOS and MEMS processes for a device, although using the capability will rely on the economics being right. "Maybe one out of ten parts would justify integration," he says.
A different outlook
The key to the timing market is finding something that resonates, and conditioning its output. Both the quartz and the MEMS providers are doing this in the physical domain. Mobius Microsystems is doing things a little differently by producing an all-CMOS chip that derives its source from an on-chip inductor-capacitor (LC) resonator. The company says this has various advantages including the ability to derive multiple frequencies and functions, lower power than a PLL, low electromagnetic interference, greater reliability than quartz, and good phase and jitter noise performance. The company is also claiming that it can deliver 90ppm frequency accuracy over temperature, process and bias variations. Tunc Cenger, director of marketing at Mobius, claims that the part is "the most accurate CMOS oscillator built, meeting the frequency accuracy requirements of the most common interface standards on the market today".
The core of the technology is a CMOS harmonic LC resonator running at gigahertz frequencies. Its output goes through a series of signal-conditioning blocks with automatic frequency centring circuitry to ensures that whatever the frequency of the resonator, the output is what has been programmed for. "We have had LC resonators in PLLs before, but we're taking it in an open-loop configuration [without a quartz crystal to lock on to] and adjusting its accuracy in real-time, using control loops on environmental variables such as temperature and voltage," says Cenger. "Our product has been embedded in a large scale ASIC that is on the market today. We are going to commercialise our product as IP and are in discussions with a very select group of semiconductor companies."
Cenger argues that the Mobius approach will offer the lowest cost; "When you do your frequency generation on chip it is the lowest costs." Hartmann of Epson is keeping a weather eye on companies such as Mobius. "We don't see them competing with us so far," he says. "We just feel that quartz is the most accurate and least temperature-sensitive material in the market. It's difficult to say if they are going to be a threat or not. We're evaluating what the options there might be but we still see the dominant part of the market being where accuracy is not that critical." Bahl at Discera is equally sceptical, citing the accuracy performance of the Mobius parts at the moment: "About five years ago the market was at 100ppm. Over the past two years the market has moved to 50ppm, and I believe the baseline of the market will soon be 30ppm.
"The LC-based approaches are very interesting and I believe that over time they will get the performance down to the right point. But the customers don't care if a device has moving parts. They have been using moving parts for decades."