Editor's Note. To view a PDF version of this article, Click Here.
The troubles facing wireless operators and their equipment suppliers have been well analyzed and documented. From overpaying for spectrum licenses and seeing subscriber growth slow, to encountering higher-than-expected costs for third-generation networks, operators have a long way to go before achieving a level of predictability and sustained growth. In the meantime, all eyes are turning to the bottom line-in particular, to the fundamentals of wireless communication-hunting for ways to maximize the capacity and coverage of current second- and 2.5-generation networks, which operators had hoped to put in the dustbin by now in favor of the much-hyped 3G upgrade.
Down the years, many techniques have been employed to increase the signal-to-noise ratio in wireless systems. However, two advanced technologies called smart antennas and multiuser detection (MUD), each of which promises anywhere from 3x to 10x capacity improvement per basestation, are now rising out of academia and the military to find a home in cellular and personal communication system networks.
Though smart antennas have been deployed in various forms for a number of years, they have been used strictly as add-ons, preventing basestations from fully realizing their benefits. For its part, MUD has yet to see a practical implementation-though trials are now being arranged. The reasons for the slow adoption are numerous, ranging from the high cost of the processing power and multiple RF front ends required (for smart antennas) to the focus on 3G networks, which hold the carrot of inherently much greater capacity-temporarily obviating the need for smart antennas and MUD.
However, technological advances along many fronts have conspired with the current state of the market to make smart antennas and MUD increasingly attractive right now. From the move to 0.13-micron processes that have delivered much more processing horsepower at a lower cost, to the advent of advanced, multicarrier RF front ends and sophisticated I/O and switched-fabric architectures that better share the processing load, it would appear that smart antennas and MUD will soon be realizing their full cost-savings potential.
How do they work?
Multiuser detection-also called joint detection, successive interference cancellation or parallel interference cancellation-augments the uplink in a CDMA network. It jointly detects all the co-users of a channel and then uses a subtractive process to cancel them out, leaving the target signal (Figure 1).
Smart-antenna systems comprise a number of antennas whose radiation patterns are controlled and adapted according to the environment by means of back-end signal processing that maximizes coverage (or throughput) while also nulling interferers. Depending on the system, the level of control can vary from basic switched antennas, which have a set of predefined patterns that might typically be reevaluated and manually changed once a month, to fully adaptive, high-end systems that dynamically alter their patterns on a user-by-user basis-even as that user moves about the sector (Figure 2).
With multiple antennas come multiple benefits, including signal gain, interference rejection, spatial diversity and power efficiency-features that lead to better range or coverage, increased capacity, multipath rejection and lower cost per basestation, respectively. However, the degree to which a smart-antenna basestation can enhance a system depends a great deal on whether the air interface uses a frequency-division or a time-division duplex (FDD or TDD) scheme.
"Most air interfaces don't have information that can be used in a closed-loop way to train the adaptive antennas," said Marc Goldburg, chief technical officer at ArrayComm Inc. (San Jose, Calif.), a leading proponent of smart-antenna systems with its IntelliCell technology. "So, the basestation directly measures the uplink and builds a picture of what it looks like. But on the downlink, it has to do that open loop-so it has to guess." According to Goldburg, that guess is based on the assumption that what goes up must come down-namely, that the downlink looks similar to the uplink, and repeats the same processing strategy that was used for the uplink's beam forming, gain points and nulls.
While that kind of guesswork might do for a TDD system, where the uplink and downlink are on the same frequency and only separated in time, it doesn't work as well for FDD systems, said Goldburg. "In FDD, the uplink/downlink channels are separated by 80 MHz at 2 GHz [in the United States], so what you find is that in some propagation environments the up- and downlinks are pretty uncorrelated." As a result, he said, the downlink performance tends to be worse in FDD systems, thereby affecting overall performance, since "interference mitigation and gain performance are determined by the weakest link."
ArrayComm has incorporated its adaptive-antenna-array technology into its own iBurst line of basestation equipment and has also licensed the technology. Currently, its IntelliCell is deployed in 100,000 basestations worldwide, with a focus on the TDD-based Personal Handyphone System in Japan and on fixed-access, wireless local-loop networks. The company also has a licensing deal with Marconi.
Capacity: FDD vs TDD
"In an FDD system, we can increase capacity by a factor of three," said Goldburg, "but with TDD, it's actually possible to reuse channels within a cell. All you need to do is provide adequate spatial separation between the users of 15 to 20 dB, depending on the air interface." With that, he said, there's no reason not to have intracell reuse of less than one, yielding up to 40 times more users per basestation.
Also leading the pack when it comes to smart antennas is Metawave Communications (Redmond, Wash.), which has taken a different road from ArrayComm by focusing on FDD networks and by manufacturing its own equipment to sell to carriers and basestation manufacturers. To date, the company's bread and butter has been its Spotlight CDMA line, a bolt-on system that essentially performs cell sculpting or sector synthesis-that is, load balancing on a per-sector basis to improve overall capacity. Customers include Nortel, Lucent and Motorola.
Nortel, meanwhile, has put its own Adaptive Antenna Beam Selection technology through trials with Sprint. Targeting the IS-95, cdma2000 1X, CDMA 1xEV-DO and 1xEV-DV air interfaces, the multicell technology trial demonstrated a twofold increase in the IS-95 voice capacity, Nortel reports.
While a bolt-on system is good for augmenting existing networks, Marty Feuerstein, senior vice president and general manager of product development at Metawave, sees maximum benefit coming from the eventual integration of smart-antenna technology directly into the basestation's signal path. To that end, the company recently scored a deal that will put its technology into Samsung's upcoming line of cdma2000 and wideband-CDMA basestations.
"Once you embed within the basestation, you can do much more sophisticated processing for CDMA-hence [obtaining] more capacity and coverage quality," said Feuerstein. Metawave's software sits right on the channel card and works on the digital baseband data in the basestation. Adaptive beam forming tracks the user with a very narrow beam. "We're tracking on a 10- to 100-millisecond basis, as they [users] move around at 60 mph," said Feuerstein.
The difference between the company's bolt-on Spotlight products and its embedded technologies has to do with placement in the signal chain, he said. "In order to do the de-spreading of the signals, you have to do it in the channel-element processing within the basestation-it's hard to do it as an add-on."
Metawave's Spotlight CDMA solution is based on an ASIC-FPGA combination, which, Feuerstein said, doesn't require "bleeding-edge" digital signal processors. "However, we'd love to take some of the processing we do now in ASICs and do that in a DSP or MIPS processor, where it becomes software configurable," he said. But that will only happen "when data rates and interference-cancellation demands rise to make the transition" from the low-cost ASIC-FPGA implementations worthwhile.
When the time comes for that changeover, there will be plenty of processing options to pick from, thanks to the move to 0.13-micron processes, said Jim Gunn, senior consultant at Forward Concepts.
At the same time, the latest DSP architectures from powerhouses such as Texas Instruments Inc., now fielding the 600-MHz C6416, and Analog Devices Inc., with its TigerSharc, exemplify what's available-and required. "For MUD or adaptive antennas you have to resort to 32-bit, floating-point processors," said Zoran Zvonar, system-development manager at Analog Devices' DSP and Systems Product Group.
Along with being able to handle 16/32-bit data in a highly parallel structure, the TigerSharc has another key feature in its arsenal, said Zvonar. "It can also handle chip-rate processing with its instruction set, hence eliminating the need for an external FPGA or ASIC-type processor." He said the TigerSharc is "the only [DSP] to do so."
The TigerSharc is also good, Zvonar said, for high-precision numerical calculations such as interference cancellation. "For that, you have to do chip-rate processing, then reconstruct and remodulate the signal. To do that, you also have to be able to efficiently operate and link different processors into an efficient processing cluster," he said. The TigerSharc accommodates this through ports that allow easy clustering. "The moral is that one engine is only as efficient as you're feeding it numbers to process," said Zvonar.
To each processor a task
But this ability to shuttle data back and forth is also a key feature of the C6416, said Sandeep Kumar, product-marketing manager at TI. "Our EMIF [Enhanced Memory Interface] supports 10 Gbits/s now," he said. With the high-performance C6416, TI finds itself better able to implement its strategy of a DSP/ASIC, fully software-configurable basestation.
"Some of our competition is trying to implement the baseband processing primarily in an ASIC, so the only time it gets to a back-end programmable unit where it can take on more functionality is when they're passing symbols back," Kumar said. At that point, he said, the user is blocked from doing any sort of weight computation or differential combining that differs from what might have been done in a normal network.
"Being able to do channel estimation and differential combining on a DSP gives people the flexibility of taking the DSP/ASIC combo and being able to do adaptive antennas," Kumar said.
The math is divided, Kumar said, between the rake receiver and search on the ASIC side, "while the DSP should have a lot of say in how the maximal-ratio combining and channel estimation take place." This setup protects the software and hardware investment, he said, while also allowing the developer to get to advanced features such as MUD in parallel or to increase the number of users.
"We're looking to incorporate things where they can do this very flexibly and then in the future be able to not even take a hit on the number of users-and still do adaptive antennas," he said. Kumar is also looking to advanced interfaces such as RapidIO for enhanced scalability as capacity demands increase.
Also homing in on RapidIO is Intrinsity (Austin, Texas), whose recently announced FastMATH processor adds a matrix-math coprocessor to the company's 2-GHz FastMIPS processor. The latter MIPS32-compliant device already comes with two RapidIO ports (Figure 3).
FastMATH for MUD
"We took a look at software workloads that need the highest absolute performance, and smart antennas and MUD definitely counted there," said Scott Gardner, vice president of sales and marketing at Intrinisty. "So, we architected a MIPS coprocessor to bolt on, in the standard MIPS programming model, to accelerate those apps."
Most of those applications are native-vector or matrix-data types, said Gardner. So FastMATH was designed as a single-instruction, multiple-data architecture "that allows us to run multiple execution engines in the math unit at the same pipeline speed as the scalar unit-that is, 2 GHz." The net result, Gardner said, is a processor capable of 32 billion multiply-accumulate operations per second. "We can do a 1,000-point fast Fourier transform at the rate of 600,000 FFTs/s," said Gardner. "Compared to the 'C6416, that's five to six times faster than TI's published numbers." Though power consumption is still to be measured, "it'll be in the power envelopes of the various embedded applications through lowering the voltage," he said.
To develop the FastMATH processor, the company worked very closely with MUD algorithm developer Mercury Computer Systems (Chelmsford, Mass.), said Veera Anantha, member of technical staff at Intrinsity.
"We've looked at Mercury's parallel-interference cancellation scheme [for MUD], and found the problem there to be, firstly, very compute intensive. The second thing is that even if you partition the problem across multiple processors, you have to move huge amounts of data between these processors, since inherently, MUD requires you to cancel the interference for each user from every other user."
Specially designed for matrix computation, the FastMATH processor is particularly suited to the parallel-interference scheme. It also addresses partitioning and scalability via its high compute power as well as its RapidIO interface. "We can support up to 48 voice users on one processor with an AMR [adaptive multirate] codec giving 12.2 kbits/s, and up to 64 users with two processors-with lots of overhead," said Anantha.
If it performs as billed, Intrinsity's FastMATH might well solve the processing problem that makes Metawave's Feuerstein balk at the idea of a fast migration to MUD. "When people talk about MUD they talk about bleeding-edge DSP farms and very high parallel processing," Feuerstein said-which pretty much would put MUD on hold for the immediate future.
But not everyone believes MUD need be so distant. "We believe it's very implementable now, and we're looking to get to trials this year," said Barry Isenstein, vice president and general manager of Mercury Computer's wireless communications group. The company has already done extensive analysis of MUD performance in the field through the PA Consulting Group (Figure 4).
Mercury finds the Intrinsity solution "very intriguing," said Isenstein. "It's one more option. We like it for its incorporation of RapidIO, since we firmly believe a [switched] fabric will be required for performance, as well as [for] the connectivity of the data and control planes."
Mercury's strategy for MUD, and for basestations in general, is to build a computer, not a MUD processor. "We see an insatiable demand for processing power as the basestation goes from fixed functions to more software-oriented plans," said Isenstein. "So we need lots of scalability." That scalability is essential, as with it comes the potential for basestations with integrated MUD, smart antennas, digital predistortion for low-cost power amplifiers and just maybe the Holy Grail of software-defined radio.
About the Author
Patrick Mannion is the Editor of Communication Systems Design. He can be reached at email@example.com.