In fact, mobile operators can end up spending as much in wireline technology as they do for wireless technologies with 4G. To address these challenges, signal compression technology offers the promise of reducing the bit rates for carrying baseband data to the radio elements, and therefore keeping fiber optic transport costs in line with existing 3G systems.

The typical remote radio head architecture for a 3G base station is illustrated in Fig 1. The radio heads and the baseband processor are different colors to illustrate that they can be provided by different vendors.

1. Remote radio head architecture for 3G wireless.
(Click this image to view a larger, more detailed version)

In the receive direction, the radio head contains all of the electronics to downconvert an RF band (typically two with antenna diversity) to an intermediate frequency, convert to digital, and then digitally downconvert individual carriers to baseband in-phase and quadrature (I/Q) pairs of samples.

Current systems utilize intermediate frequencies in the range of 120 to 180 MHz, and employ dual 14-bit ADCs sampling at 122.88 MHz. Likewise, in the transmit direction, I/Q baseband samples are modulated onto a digital IF carrier which then passes through processing to reduce the crest factor, and to digitally predistort the signal to compensate for sidelobe generation in the power amplifier. Since this digital processing occurs primarily in FPGAs, this market has become one of the most important for the FPGA vendors.

The baseband samples (I/Q) in both the receive and transmit directions are multiplexed onto a high speed SERDES interface using either the CPRI or OBSAI standards. CPRI intellectual property is available off the shelf for leading FPGA families. The radio head sends the serialized data stream to the baseband processor over a fiber optic cable.

If the baseband processor is located at the tower site, the length of the cable may be modest, a couple of hundred feet. However, often times for residential coverage or coverage along highways, the radio heads may be daisy-chained and thus be remote from the baseband processor. The bit rate over this fiber optic connection directly affects installation costs as well as operating costs if the mobile operator must lease fiber optic facilities.

If the raw data from the ADC were transported over the fiber optic cable, the data rate would be 4300.8 Mbps = (2 antennas × 122.88 Msample/sec × 14 bits/sample / (8B/10B) = 4300.8 Mbps), which exceeds the capacity of low cost fiber optic transceiver modules developed for the storage area networking market, particularly 4G Fibrechannel.

Consequently, channelization and decimation of the data is required. For wideband CDMA, the received carriers are typically decimated to twice the chip rate of 3.84 MHz to a sample rate of 7.68 Msample/sec. Since W-CDMA receptions from handsets are power controlled to low signal to noise ratios, very few bits per carrier need to be transported across the fiber optic interface with narrowband digital AGC applied at the output of the channel filters.

Thus, for a 20 MHz radio with four W-CDMA carriers, the aggregate bit rate is a modest 614.4 Mbps = (2 antennas × 4 carriers/antenna × 7.68 Msample/sec/carrier × 4 bits/sample × 2 (I/Q) / (8B/10B)). This is the predominate bit rate for 3G remote radio heads and can easily utilize low cost SFP fiber optic transceiver modules developed for the OC-12 and gigabit networking market. Three such radio heads can also be daisy-chained within the capacity of an OC-48 (2.5 Gbps) link, enabling a single fiber to be used for all three sectors on a tower, or for coverage along highways or in rural areas.

Fourth generation wireless systems such as LTE (Long-Term Evolution) will require four antennas per sector in both the transmit and receive directions to support multiple-input-multiple-output (MIMO) baseband technology. In addition, WiMAX systems can support smart antenna beamforming technology with up to eight antennas per base station.

Furthermore, for MIMO and beamforming technologies, narrowband AGC cannot be applied in the radio head without interfering with the antenna-combining function in the baseband processor, so often the full dynamic range of the signal must be preserved. Utilizing the same ADC configuration, but now allowing for bit growth through the digital downconverter and channel filter, the resulting signal can contain 16 bits of dynamic range.

With OFDM signals, the channel filter must allow for decimation to the FFT sample rate which is defined as the channel bandwidth. With the same 20 MHz of bandwidth divided into perhaps two 10 MHz channels, the bit rate over the fiber optic cable now becomes 3.2 Gbps = (4 antennas × 2 carriers/antenna × 10 Msample/sec/carrier × 16 bits/sample × 2 (I/Q) / (8B/10B)).

At 3.2 Gbps, the bit rate exceeds OC-48 capacity driving the cost for long haul fiber transceivers. This bit rate also requires the high-speed SERDES interfaces of high-end FPGAs. Moreover, for an 8 antenna beamforming system, the fiber bit rate is doubled to 6.4 Gbps, which now requires very expensive 10 gigabit networking transceivers and FPGA SERDES interfaces.

By employing signal compression, the fiber optic bit rates can be reduced to enable the continued use of low cost fiber optic transceivers. For example, with a compression ratio of 1.28:1, the 3.2 Gbps of data from a 4×4 MIMO LTE can now fit into an OC-48 link at 2.5 Gbps.

One such signal compression algorithm is Samplify's Prism, which provides lossless and near lossless compression for a wide range of wireless signals including W-CDMA and orthogonal-frequency-division-multiplexing (OFDM) waveforms. Prism has a mode called RateTrak, which adapts the dynamic range in the signal to maintain a user-specified compressed bit rate.

Fig 2 shows the results of applying Samplify's Prism compression algorithm in RateTrak mode at a compression ratio of 1.35:1. The decompressed signal is indistinguishable from the original waveform, and the distortion introduced by compression is 64 dB below the signal peak, and spectrally flat, except around the unused subcarriers at DC.

2. 10 MHz channel spacing OFDM signal compressed at 1:35:1; Full-scale (top) and 12 dB back-off (bottom).

RateTrak will adapt the dynamic range of the signal to meet the output bit rate requirements. In the case where the input signal level is below full scale, RateTrak will automatically preserve more of the least significant bits, as illustrated by the lower plot in Fig 2, where the signal level has been lowered by 12 dB, yet the distortion level introduced by compression is still 60.4 dB below the signal.

Note that RateTrak does not affect the sample rate of the signal during compression and decompression, so the FTT function of the OFDM demodulator can operate directly on the decompressed samples. In contrast, to achieve 35% reduction in the bit rate, one could simply quantize the four least-significant bits from the 16 bit signal resulting in a 12-bit signal. The equivalent noise floor for quantization would be at 24 dB in the graphs in Fig 2. The resulting signal to quantization noise level for the full scale signal would be only 54 dB. This result is 10 dB worse than the signal compression case.

With quantization, for every 1 dB decrease in the signal level, the signal to quantization noise level also decreases by 1 dB. This is not the case for compression, where a 12 dB decrease in the signal level, led to only a 3.5 dB decrease in the signal to distortion level. At this signal level compression outperformed quantization by over 18 dB in noise level. Table 1 summarizes these results.

Table 1. Compression versus Quantization.

The distortion introduced by compression will be limited by the error-vector magnitude limit for OFDM waveforms. The worst case is for 64 point quadrature amplitude modulation (64QAM) on each OFDM subcarrier. This limit is 3.1%, or –30.2 dB for WiMAX per IEEE 802.16d.

The rate-distortion curve for Samplify's Prism signal compression is illustrated in Fig 3. For compression ratios of 2.4:1 or lower, the distortion introduced by compression is less than the EVM limit for 64QAM. At compression ratios of 2.1:1 or less, the distortion introduced by compression 10 dB below the EVM limit.

3. Rate-distortion curve for 4G Wireless.
(Click this image to view a larger, more detailed version)

The foregoing discussion describes how signal compression can be used to achieve bit rate reductions for 4G wireless baseband signals to enable them to into cost effective fiber optic technologies.

Using the CPRI and OBSAI standards places additional constraints on the choices of bit rates which must be used. Since the CPRI specification was originally conceived for W-CDMA with a 3.84 MHz chip rate, the possible line rates specified in the standard are all multiples of this fundamental frequency.

Although a 3.84 Mchip/sec W-CDMA signal has a channel spacing of 5 MHz, all W-CDMA demodulators operate at a samples rates of equal to or twice the chip rate. Obviously, this relationship between the chip rate and the CPRI line rate also makes for simple decimation in the digital down converter, and time-division multiplexing of the I/Q outputs of the DDC into a single serial stream.

However, as mentioned earlier, for OFDM waveforms, the channel bandwidth is the same as the FFT sample rate. Hence for the same 5 MHz channel bandwidth, and LTE signal will require a different sample rate than a W-CDMA signal in the same channel. To map LTE onto CPRI, a complex sample rate conversion algorithm must be employed at both ends of the fiber optic link. For FPGA implementations, these sample rate conversion algorithms can be a complex as the original digital down conversion function.

With WiMAX, the mapping from OFDM sample rate to CPRI line rate is far more complicated due to the number of different channelization for WiMAX used around the world. WiMAX deployments are planned in the 2.5 GHz band in North America where legacy 6 MHz TV channels were once employed. In Europe and Latin America the 3.5 GHz and 3.8 GHz bands employ 7 MHz channelization. These are in addition to the 5 MHz used for WiBro in South Korea, and other allocations in the 2.3 GHz band. Each of these different channelizations must be mapped onto CPRI line rates which are a multiple of 3.84 MHz.

While IP exists for the CPRI framing function in leading FPGA families, the problem of mapping the different channelizations into this framing structure is left to the system OEMs. One WiMAX system OEM reported taking 6 man months of effort to perform just one of these mappings.

Instead, signal compression can be used to perform the rate adaptation with significantly lower complexity compared to resampling filter structures. Since all channelizations can be supported by changing the compression ratio parameter, the mapping needs only to be designed one-time, rather than 6 man months for each channelization required.

Table 2 details the compression ratios used to map the worldwide channelizations for WiMAX and LTE onto CPRI antenna-carrier containers. With up to 2.08:1 compression, all channelizations up to 7 MHz for a 4 antenna MIMO system can be carried in the same 614.4 Mbps CPRI line used in existing 3G systems.

For 10 and 11.2 MHz channelizations with up to 1.46:1 compression, the 1228.8 Mbps CPRI line rate may be employed, staying within gigabit networking class fiber optics, rather than having to move to the more expensive OC-48 class transceivers.

Table 2. Compression settings for mapping LTE and WiMAX onto CPRI.
(Click this image to view a larger, more detailed version)

Signal compression and decompression must be implemented at each end of the fiber optic link. Even without compression, system designers are choosing FPGAs for implementation of the CPRI multiplexing and demultiplexing function and the sample rate conversion function required to support OFDM signals. Baseband processing might then be performed in a DSP or ASSP. These FPGAs can also host the signal compression and decompression functions.

In addition, there exists an analog-to-digital converter which has integrated signal compression, the SAM1610 from Samplify. A single SAM1610 provides 16 analog input channels each at 12 bits of resolution, enough for processing I/Q baseband signals from each of eight antennas, as would be the case for an WiMAX beamforming base station.

A sister device, the SAM1605 provides eight analog input channels, enough for I/Q baseband signals for a 4 antenna MIMO sector. Both devices include Samplify's Prism signal compression, and support RateTrak mode operation. Optimized for I/Q baseband sampling, these devices also provide other system cost benefits by enabling low cost direct conversion radio architectures, and replacing two or four expensive high-IF sampling ADCs.

Allan Evans Bio/contact information Allan Evans is the VP of Marketing at Samplify Systems, Inc., a fabless mixed-signal semiconductor company that combines high performance analog with sophisticated digital processing to deliver a new class of intelligent data converters.

Samplify provides the only real-time lossless and lossy compression solutions for high-speed sampled data systems. Delivering "simply the bits that matter," Samplify's Prism compression technology turbo-charges the I/O subsystem of Samplify data converters, reducing bandwidth and storage bottlenecks in DSP systems without the power, area, and cost of brute-force hardware solutions.

Allan holds an MSEE from UC San Diego and an MBA from Santa Clara University. He can be reached at aevans@samplify.com.