The evolution of SERDES designs for cellphones

Slimmer cellphones with more functions and features have created a number of system-design challenges. The problems are well-known: managing energy consumption and minimizing noise while moving data around in a handset. The situation provides an opportunity for silicon designers to tackle a system-level-design problem. By applying experience designing SERDES (serializers/deserializers) for devices from laptop computers to cellphones, Fairchild’s design team learned valuable lessons about the systems approach.
A product initially targeting cellphones required an in-depth study of the systems and in-depth discussions with several customers. What Fairchild now calls its µSERDES (micro-SERDES) family had design requirements that differed greatly from versions developed previously for computers. Differences between the previous SERDES and the cell-phone application included issues that the Fairchild team faced for the first time, such as serialization of a microcontroller interface, power as a critical design factor, the need for bidirectionality, and the fact that a radio would be an integral part of the application using the SERDES.

The needs and perceptions in this market differed so greatly from those in other applications that the semiconductor supplier and systems manufacturers even used different terms. The systems manufacturers worried about ESD (electrostatic-discharge) events that would temporarily cause the LCD to stop working, and the semiconductor manufacturers worried about ESD events that would permanently destroy the device. The difference in views over this one small question of ESD was sufficiently interesting to warrant the summary that Table 1 provides. Needless to say, all the first-time technical issues and supplier-user nomenclature and perspective differences made for several surprises along the way.

IDENTIFYING THE OPPORTUNITY
It all began at a customer trade show in 2001, when a product engineer from a cell-phone manufacturer and one of Fairchild’s engineers began discussing the flip, or clamshell, phone design. The application intrigued the Fairchild engineer, who had been involved with the novel implementation of SERDES for laptop computers. In many ways, the situations were the same: The displays were in different places from the processors. In the computer application, SERDES technology allowed the application to pass government EMI (electromagnetic-interference) testing, such as FCC (Federal Communications Commission) requirements. The increasingly higher frequency interfaces were using LVCMOS (low-voltage-CMOS) signaling that created a very-high-EMI environment that could no longer meet the EMI limits. With this previous experience, the Fairchild engineer concluded that a SERDES approach could be as important for the cell-phone applications as it had been for laptop computers. As it turned out, the EMI issue was a much larger concern than for the laptop from a systems perspective. However, it was also a much smaller concern, when selecting devices, than two other key issues: susceptibility and the number of wires.

USER INPUT
As the Fairchild design team became more focused on providing a SERDES option for cellphones, its members realized that an active engagement with the key cell-phone customers’ design engineers would be crucial to success. So, the team traveled the world to discuss the application requirements with every major cell-phone manufacturer. It soon became obvious that this application was not standard for a SERDES. The cell-phone SERDES would require some radical changes to be acceptable.

After the initial discussions with customers, the team found that the system may not always be synchronous and that there may not be a defined free-running clock to load and transfer the data. An important design consideration was a bidirectional datapath, which was required to serialize a microcontroller datapath; a clock, synchronous with the data, would not be available. In addition, write as well as read cycles would be beneficial.

These initial requirements indicated a radically different implementation of a SERDES and turned out to be just the initial design surprises. The team would need to decouple the parallel-input interface from the serial interface as far as timing was concerned. Data would not be coming at convenient, regular clock intervals. Data would arrive whenever the microcontroller “felt like” sending it. For a read operation, while data was going in one direction, which happened to be the wrong direction, control signals still needed to go in the opposite direction. Both of these requirements were well beyond the capability of a standard run-of-the-mill SERDES.

A high-end SERDES, such as an Ethernet SERDES, had some of these qualities but was far too expensive for this type of application and did not match all of the requirements. Packetizing the data, including header information, error detection, and error correction for transferring information to a display approximately 10 cm away was clearly overkill. The design needed to be a basic SERDES that could transfer data in both directions and use an asynchronous signal to latch the incoming data.

Another key insight—another design surprise—occurred at this time. Although previous SERDES-application designers had to worry about FCC regulations, this one had much worse issues to deal with concerning EMI. Previous applications had to be below radiation levels of –60 dBm to make sure the system did not interfere with other systems. The cellphone could not tolerate EMI at such levels, due to the radio. The µSERDES design could not interfere with radio transmission and reception from an antenna that may be only a quarter of an inch away. So, –60 dBm was no longer the target; –120 dBm was the new target.

The next specification that was out of line was power consumption. The power consumption of the typical SERDES was off by orders of magnitude from the requirement. Battery life is one of the most important factors in cell-phone design. Every milliamp of operational current is a concern for talk time, and every microamp of current is a concern for standby time. As the team dissected SERDES implementations, it became clear that there were three main components of power consumption in the SERDES pair. Both the digital logic in the device and the serial link would consume significant current, as would the PLLs (phase-locked loops). The team had determined that it would use differential signaling for this application due to the minimum radiation it provided, as well as its high bandwidth. At this point, the engineers had not realized the importance of this decision. For the most significant reduction in power, the PLLs needed further investigation. Both the operational current and the standby current were unacceptable. Standby current was off by three orders of magnitude.

A PRELIMINARY SPEC
All of this information was beginning to take form in an initial product specification. The engineers needed a bidirectional device with serialization for at least a camera and an LCD, current consumption of less than 5 µA at standby and less than 7 mA in operation per device, and less than –120-dBm EMI. Another key item for this market was that most of the engineers who would use the device had never dealt with serialization, so it was also critical that they make the device as easy to use as possible.

A rough sketch of a device showed how it would operate and meet most of the criteria. However, a few questions still remained concerning how to provide the key features. Figure 1 shows the approach the engineers began presenting to customers.

The figure depicts a design close to the device that went into production for the initial generation. It is bidirectional and can work with a microcontroller interface or a pixel interface. Of course, some questions still remained about how to hit some of the parameters, such as power, EMI, and microcontroller-read operation, as well as the separation of parallel and serial clocking. For example, EMI needed to be much lower than previous SERDES with a target of –120 dBm versus a typical target of –70 dBm. It was also necessary to provide an approach for a standard pixel interface with pixel clock, vertical synchronization, horizontal synchronization, and pixel data. But that exercise was rather straightforward, and the team expected it to be similar to the previous SERDES.

Now, the engineers had some important insights into the application and the customers’ needs. It was clear that every customer required both pixel and microcontroller interfaces that varied according to platform needs. In some applications, the baseband processor required a microcontroller interface to the LCD, whereas other baseband processors required the more standard pixel interface. Because almost every system manufacturer used processors from multiple suppliers, a single manufacturer almost always required both interfaces.

So, it would be advantageous to provide a single device that could provide an approach for both interfaces. In addition, the same device needed to provide both the parallel-to-serial and the serial-to-parallel transition. This approach would be a huge advantage to customers, because they could purchase one product rather than four or six devices. It would reduce inventory requirements and provide a volume-pricing advantage. Design decisions specified that the device would be bidirectional, with each device containing both a serializer and a deserializer for microcontroller operation, and that the serial I/O would be differential to reduce EMI.

One of the critical breakthroughs for the design team involved the differential I/O. All differential I/Os, although called current mode, used voltage-level receivers. The receivers looked for a voltage difference between the two signals to determine a one or a zero. After careful consideration, the design team decided to look instead for current direction. Like a lot of good ideas, the advantages seem obvious in retrospect. Current direction is more stable than voltage levels, so the amount of drive necessary to distinguish between a one and a zero should be significantly less. In fact, it was possible to go as low as a 50-mV swing rather than the minimum of approximately 250 mV of voltage differential. That value represents five times less drive across the same 100Ω termination. The team termed this new I/O technology CTL (current-transfer logic).

CONFIRMATION/FEEDBACK
The new technique required another round of customer visits to explain this concept and seek further insight into the application requirements. In discussions with customers on this round, the team gathered a few more key details. The camera and the LCD were in a similar location, and the datapaths on both ends were in a similar location. The engineers had a bidirectional µSERDES, so why not try to use a single pair to perform both sets of operations? One of the most critical pieces of information, essentially another design surprise, came to light at this time.

A knowledgeable engineer at one of the key customers to which we were presenting CTL and the advantages of the low swing for EMI and power consumption mentioned that the low swing could be a problem due to susceptibility. “Susceptibility” is the electrical signal’s vulnerability to radio noise that could disturb the signal enough to cause data corruption. This insight was absolutely key. Now, there were barriers on both sides of the signaling equation. On one side, the team had a radiation issue, so it needed to make the signal as small as possible. On the other side, the team had radio noise that could cause significant disturbance to the signal if it were too small.

The engineers were getting close to a reasonable solution, and the customers were intrigued with the idea of using a single pair of µSERDES to even further reduce the number of wires on the flex without adding another pair of devices. An interesting point here is that it was becoming clear that the customers had little fear of EMI. They had always had to deal with it, so EMI was less of an issue than the number of wires. Reducing the number of wires was the struggle for system designers due to the advent of the flip phone and the need to send all of these signals through a tiny hinge between the flip and the base of the phone. They could have gone with LPLVDS (low-power low-voltage differential signaling) as the easier path, but they had great confidence in CTL. There would be considerable advantages of supplying a signaling structure that would emanate less than –120 dBm, because it would reduce the need for shielding on the flex for the differential signals. More important, the shielding could affect the characteristic impedance of the flex. In addition, CTL could provide power savings and is also robust for susceptibility due to the receiver structure.

FINAL DESIGN DECISIONS
The approach for the microcontroller-interface clocking was to completely separate the timing of the serial and the parallel clock. The team still needed to generate the serial clock through PLL-based multiplication and the main control signal—for example, write enable—to strobe in the parallel data. The extrapolation of this insight would lead to the development of a second-generation device, ULP (ultralow-power) µSERDES.

On the power front, CTL would significantly reduce power over LPLVDS, but those savings were not enough. The team needed to eliminate the use of at least one of the PLLs, which would provide a significant power savings. Between eliminating one of the PLLs and reducing current with the use of CTL, the team could almost halve the current. The problem was that the team needed to send three sets of information between the serializer and the deserializer: the data, the bit clock to clock the individual bits as they arrived at the deserializer, and the word clock or word boundary that identified the first bit of a word for the deserializer.

Unless the engineers were willing to send three signals over six wires, they would need a PLL at the deserializer to provide either the bit clock or the word clock. Otherwise, they would have to come up with a seriously innovative approach. They concluded that it was critical to discard one of the PLLs, and the only place to eliminate one without placing an unreasonable burden on the customers was at the deserializer. The simplest approach was to send the serial data along with a bit clock to clock the data into the deserializer. Now, the challenge was how to send the word clock without using a PLL or adding another differential signal.

One of the engineers determined that the team could solve the problem if it could do something clever with the two signals to encode the word boundary. Adding two bits to the serial stream produces a unique pattern to define a word boundary. At the point in time of these two bits, if the clock did not toggle but the data did, a singular sequence would occur that you could easily identify as the word boundary. Even better, you could identify it exclusively for every word. Most schemes for encoding a clock take a PLL and thousands of words to find the word boundary through a repetitive recognition algorithm. This ingenious scheme allowed immediate recognition of the word boundary. If the system lost the word boundary due to a huge noise event, it would lose only one word.

DESIGN TRADE-OFFS
The team had now knocked down just about every known barrier to provide an initial SERDES approach for cell-phone designers. The units satisfied the needs of the customers and provided state-of-the-art power, EMI, and susceptibility, and they were usable with a microcontroller or a pixel interface.

The engineers needed to make a number of trade-offs on this first-generation product (Figure 2). Table 2 lists some of the most important ones. The main parameters involved power, size, and cost—all of which were equally important for the cell-phone design.

The major disappointments on this project were that the half-duplex approach that the team originally envisioned for using a single pair of µSERDES was just too cumbersome to implement at a system level, and the read operation did not work as well as the team had expected.

Because of the systems approach and close interaction with customers in the design phase, the company started shipping products as soon as it received approval for production. The team was happy with the success of this project. There were three variations to accommodate the camera and LCD, as well as a product for special applications. A 12-bit device targeted the camera, which is almost always an 8-bit-pixel interface. This device has also seen use in 8-bit-LCD-pixel interfaces. A second version targeted the typical 16- or 18-bit-LCD-pixel interface. A third part targeted microcontroller usage.

Because the SERDES concept was so new to cell-phone designers, the engineers developed detailed application notes on the modes of operation and provided evaluation cards to ease testing. To further simplify the implementation of technology that was a significant departure from previous cell-phone designs and because schedules were so critical, they provided cell-phone hookups showing the use of the µSERDES in the applications. The applications group acquired multiple cellphones from customers, took them apart, inserted the µSERDES pair into the designs, and demonstrated that they worked.

The next step
In parallel with the initial project, the team had begun identifying the next-generation product. This next product improved the initial design concept but required development time that was outside the scope of the initial project. Recently introduced ULP devices use the separation of the parallel and the serial clocking for even greater reduction in power consumption.

Because there was no need for synchronizing the parallel and serial data streams, it was possible to eliminate the second PLL. Eliminating all PLLs had been a goal since the beginning of the first-generation product. The team realized that it could use just an internally generated clock to provide the clocking for the serial stream. It significantly lowers power and provides a microcontroller interface that does not require a system clock.

Eliminating the second PLL was a breakthrough on its own. It turned out to be even better when the team realized that the clock could always run at the maximum frequency of the specifications. In this case, the serial stream could burst across the interface, and the pair of devices could go into a pseudo-power-down state, burst standby, for most of each transfer cycle instead of staying powered up continuously.

Engineers often regard dynamic power as dependent on the frequency of operation. This assumption is inaccurate. The reality is that dynamic power depends on the number of edges per unit of time. Because the team was transferring a specific number of bits, or edges, per unit of time, the dynamic power did not suffer due to the higher speed. Instead, the burst standby after the transfer considerably reduced the static power. Furthermore, for a microcontroller that is not in continuous operation, the average power would drop drastically due to the burst standby between operations.

As a comparison, the first-generation 24-bit FIN24xx had state-of-the-art power consumption, using roughly 14 mA per pair of devices for a 5.44-MHz pixel interface with a constant clock. The 24-bit FIN324, a ULP device, consumes roughly 6 mA per pair of devices for a 5.44-MHz pixel interface (Figure 3). A microcontroller interface requires an assumption of percentage of time of operation. A reasonable estimate is that the microcontroller operates roughly 30% of the time. This estimate gives a power number of approximately 3 mA if the microcontroller cycle frequency is the same 5.44 MHz.

In the end, the engineers’ experience showed them that you can apply expertise in SERDES design from one market—the computer industry—to a different market with different requirements. But making this transfer required intensive research, dialogue with customers, and a willingness to shift the focus of design from the chip to the end system (Figure 4). It also meant accepting that design communities have not only different concerns, but also different language for discussing them.

AUTHOR INFORMATION
Michael Fowler is a member of the technical staff and a member of the µSERDES design team of Fairchild Semiconductor. You can reach him at michael.fowler@fairchildsemi.com.


Click here for Illustrations:

Figure 1

Figure 2

Figure 3

Figure 4

Table 1

Table 2