Thursday, December 7, 2023

Radio Architecture Matters: A Review of RF Sampling vs. Zero-IF

Radio Architecture Matters: A Review of RF Sampling vs. Zero-IF

by Brad Brannon Download PDF

Abstract

The growing demand for wireless services not only challenges our limited spectral resources, it also challenges the radio designer to select the correct radio architecture. A proper radio architecture not only provides solid performance, but it simplifies the circuit around the radio to minimize cost, power, and size. In the era of increasing radio deployments, a proper radio is tolerant of current and future wireless neighbors that might otherwise cause interference down the road. This article will examine two common radio architectures and compare the trade-offs on how each solves the unique challenge of growing co-location issues.

A Growing Challenge—New Wireless Neighbors

When the wireless revolution began some 30 years ago, there were only a handful of bands—which were confined mostly below 900 MHz—and typically there was one band per country. As demand for wireless services grew, new bands were steadily added and now there are 49 bands1 globally assigned to 5G NR alone, not counting mmW allocations. Most of the newer spectrum is above 2.1 GHz with bands covering 500 MHz (n78), 775 MHz (n46), 900 MHz (n77), and as much as 1200 MHz (n96).

As these new bands come online, one of the biggest challenges is how to ensure adequate receiver performance in the presence of blockers in these legacy bands. This comes mainly from the co-location requirements where they are deployed, with bands 2, 4, and 7 in the U.S., and their counterparts, bands 1 and 3, in other regions. This is particularly critical for wideband radios servicing applications in n48 (CBRS) and any portion of n77 or n78.

Wireless demands will continue to grow in the future, and the challenges with co-location and interference are always present.


Radio Designs and RF Protection and Selectivity


One of the key challenges for a receiver design is protection from signals that are not of interest. From the beginning, radio engineers have sought different ways to accomplish this, initially with brute force filtering and later with various heterodyning techniques with distributed filtering. Over the years, three key architectures have been developed to solve these challenges: direct conversion (zero-IF), super-heterodyne (IF), and direct RF sampling. While IF sampling is popular, it will not be the focus of this article. Instead, the focus will be on comparing RF sampling and zero-IF since these are currently the most progressive implementations in the wireless domain. Each technique introduces different engineering trade-offs and varying impact on the surrounding circuits and their requirements. This includes the method of frequency translation, the amount of RF and baseband gain, how RF images are dealt with, and how and where filtering is implemented. Details of these trade-offs are shown in Table 2.

Gain Distribution and Power Dissipation

RF sampling and zero-IF have key differences in their gain distribution. As shown in Figure 2, RF sampling has all the gain in the RF domain since all frequencies in the radio remain constant as the signal is processed. For comparison, Figure 1 shows a zero-IF architecture. For this architecture, part of the gain is at the RF frequency, but the balance is at baseband after the frequency translation.

Figure 1. A typical zero-IF signal chain.

Figure 1. A typical zero-IF signal chain.

Figure 2. A typical RF sampling signal chain.

Figure 2. A typical RF sampling signal chain.

There are trade-offs to either architecture. From a gain perspective, gain at higher frequencies requires more DC than lower frequencies due to the higher slew rates required, especially as the signals get progressively larger within the signal chain. This means that an RF sampling architecture dissipates more power in the linear RF section than does a zero-IF, where a significant portion of the gain is at DC. At lower frequencies, the slew rates are lower and thus standing currents can be correspondingly less.

The challenge with RF sampling is the requirement to drive a largely capacitive input (sampling capacitor) at both high frequency and at relatively high voltage (~1 V). By contrast, a zero-IF input is a well-behaved 50 Ω (or 100 Ω) into a summing node of a baseband amplifier, which is providing gain, eliminating and isolating the sample node from the RF signal, and reducing the RF drive required by the gain provided. This has a profound impact on power consumed in the linear RF section, as it reduces the total RF dissipation between 25% and 50% in favor of zero-IF architectures by eliminating a third RF gain stage and the lower standing currents required for baseband vs. RF amplification.

In addition to linear power is the power associated with the digitizer. With zero-IF converters, only the bandwidth required is digitized. With RF sampling, not only is a wide RF bandwidth digitized, but the sample rate far exceeds the Nyquist requirements. Both bandwidth and sample rate are expensive in terms of power. Exact power depends on the process, but when implemented on the same process, RF converters consume about 125% more power than baseband converters for a typical single band application. Even where two bands might be digitized by an RF converter, the power penalty is still more than 40%.

Table 1. Gain Distribution in Different Architectures

Zero-IF RF Sampling
RF Gain 32 dB ~50 dB
Baseband Gain ~18 dB

Images and Spurious Signals

There are secondary trade-offs in these options as well. For example, zero-IF introduces LO leakage and I/Q mismatch image terms,2 while RF sampling introduces interleave spurs3 due to mismatches within the converter architecture, as well as RF harmonics in the converter and sample-related jitter terms.4 The good news is that most images and spurious signals are mitigated with various background algorithms regardless of architecture.

These two architectures have vastly different frequency plans that impact how aliasing is handled and how much RF (external) filtering must be applied. Aside from architectural spurious signals, all radios will generate RF harmonics and are subject to aliasing.5 RF sampling radios take advantage of aliasing to downconvert the desired signal if it is naturally beyond the first Nyquist zone. However, it is the response of unwanted signals that are generally the issue given that they may fall on top of desired signals inadvertently after they have aliased. These signals must be mitigated by careful frequency planning, by aggressive RF filtering, or by high enough sample rates that there are no aliases. Each of these comes with challenging trade-offs.

Zero-IF architectures translate the signal to baseband (near DC). While RF harmonics are certainly created, they mix well away from the baseband in all cases and are adequately filtered by the low-pass response of the typical zero-IF input structure noted in following paragraphs. Similarly, aliasing is also mitigated by the relatively high sample rates of the baseband sampler used and the selfsame input structure.

Zero-IF Filter Requirements

One easily overlooked feature of a zero-IF architecture is that the baseband input amplifier is typically constructed as an active low-pass filter that operates as an integrated analog filter, which greatly reduces the analog filter burden. In conjunction with on-chip decimation filtering, it also functions as a programmable channel filter to eliminate signals closer than those associated with Nyquist. In addition, the sampling devices within zero-IF receivers typically include feedback that provides additional out-of-band rejection. In effect, this means that the out-of-band regions of the radio have a larger full-scale range than the in-band. As demonstrated in previous writings6 and simplified in Figure 3, zero-IF radios inherently have a good tolerance to out-of-band signals. In Figure 3, the vertical axis represents the input power level relative to in-band that would cause a 3 dB desense, showing that in-band signals have a built-in tolerance to out-of-band signals not found in other architectures.

Figure 3. An example of the impact of on-chip zero-IF filtering.

Figure 3. An example of the impact of on-chip zero-IF filtering.

Because of this built-in filtering, the primary concern becomes protection of the RF front end—that is, the LNA. A typical configuration will include a SAW filter between the first and second stage LNA for FDD and some TDD. Some TDD applications will have the SAW filter after the second stage, but the second stage is bypassable under large input conditions, as is shown in Figure 1. Typically SAW filters will provide about 25 dB of out-of-band rejection, and that is assumed here. In addition to the SAW filter, a cavity filter is required on the antenna side of the LNA, which is shared with the transmitter.

A typical LNA might have an input 1 dB compression point of –12 dBm. If the out-of-band or co-location requirements are 16 dBm, these signals must be filtered to about 10 dB (or more) below the input 1 dB compression point of the LNA. This is a minimum of 38 dB rejection (+16 – –12 + 10). Including the SAW filter, this is a total out-of-band rejection of 63 dB as presented to the input of the zero-IF. Assuming RF gain does not roll off, and including the total filter rejection up to the core radio input, the maximum out-of-band signal level will be –20 dBm. This is well below the typical full scale and will be further attenuated by the on-chip filtering previously explained. No spurious signals or desense would be anticipated from this input level when compared to Figure 3.

RF Sampling Filter Requirements

There are two concerns when working with RF converters that require direct attention for filtering. First, any signal regardless of input level can create undesired spurious signals that can occupy the same frequency as the desired signal. Interleave related spurs are dealt with by algorithms, but architectural spurs are another issue as they can be unpredictable. For many older RF converters, this was a constant challenge to radio performance. Fortunately, many new converters include background dither7 in one form or another to mitigate these issues and present relatively clean SFDR sweeps, as shown in Figure 4.

Figure 4. An example of a converter with dither.

Figure 4. An example of a converter with dither.

Table 2. Engineering Trade-Offs Between Architectures

Zero-IF RF Sampling
Overall Architecture Pro: Easily implemented in a frequency agile radio in a low power monolithic single-chip design. Con: Channel bandwidth will be limited by baseband bandwidth. Pro: Very wideband radios can be implemented. Con: Relatively high power solutions and requires discrete external filtering for all selectivity.
Frequency Translation Quadrature demodulator Sample cap and digitizer
Pros: inherent alias protection, low power Cons: LO leakage, baseband images Pro: simple digitizer implementation Cons: high power, prone to aliasing, jitter/phase noise4
Gain RF: ~32 dB
Baseband: ~18 dB
RF: ~50 dB
Baseband: N/A
Pros: lower total dissipation, baseband gain easily integrated along with active filtering, input impedance easily managed Con: bandwidth limited by amp Pro: very wideband radios attainable Cons: high OIP3 drive amplifier required (high power); input impedance typically capacitive unless high power buffer used
Images LO leakage, I/Q imbalance, baseband harmonics Direct aliases, Interleave artifacts, RF converter harmonics
Pro: RF harmonics and converter aliases fall out of band Con: subject to LO leakage, I/Q imbalance (can be fixed with algorithms) Pro: no LO leakage or I/Q imbalance terms Cons: interleaved spurs (fixed with algorithms), subject to aliases, subject to RF harmonics and clock-related phase noise
Filtering Distributed between RF and baseband Single frequency
Pros: integrated alias protection, limited external filtering required due to filter integration Cons: none known Pro: requirements are easy to derive Con: high complexity filter required

What is notable in this SFDR vs. input level plot is that the first 15 dB show degradation due to slew rate limitations in the converter, which will typically generate strong second and third harmonics that must be abated. Once the RF input is below this level, harmonics and architectural spurs are typically no longer an issue (consult converter performance to verify). With a full scale of 1 dBm, it can be expected that spurious signals will reduce significantly by the time out-of-band signals are rejected below –14 dBm into the converter. With a conversion gain of 50 dB, as shown in Table 2, this equates to –64 dBm at the antenna. If the input is potentially 16 dBm, then the RF filtering needs to be 80 dB or more for non-aliased cases. Assuming a SAW filter provides 25 dB, this leaves 55 dB for the cavity filter to adequately protect the RF ADC from generating nonlinearities due to out-of-band signals as well as protect the input of the first stage LNA from being driven into nonlinearity by out-of-band signals. This example represents a well-behaved converter, but the SFDR vs. input level of the converter selected should be closely examined to determine if more filtering is required.

There is one additional concern for RF converter architectures based on current merchant silicon, and that is alias protection. Current RF converters are based on cores that operate between 3 GSPS and 6 GSPS. At these low rates, it is impossible to avoid aliased terms without the use of aggressive filtering to mitigate the impact of aliasing. This problem only abates after sample rates reach double-digit GHz.

A simplified way to consider the impact of aliasing on the filter requirements is to consider the protection of a single resource element from the aliased 16 dBm co-location requirement. The goal is to suppress the aggressor to the point that should it alias to a desired RB, it is filtered sufficiently that no disruption occurs. A wide area reference channel based upon a G-FR1-A1-4 signal would account for a signal level per RB of –118.6 dBm at approximately 0 dB SNR. Therefore, the offender must be filtered to 10 dB to 15 dB lower, or about –130 dBm, to prevent disruption. Thus, a total rejection of about 150 dB is required, or about 125 dB from the cavity filter with one SAW filter providing the balance of filtering.

Filter Summary

Figure 5 shows the cavity filter requirements for both RF sampling and zero-IF. Because the RF sampling architecture has two separate requirements, the most restricting will dominate and a realizable filter would simply have to meet the most restrictive or 125 dB rejection to cover the entire band. While this filtering is readily available, it comes at the cost of a bulky filter. In contrast to the zero-IF architecture, where only 40 dB rejection is required, the result is a significant weight and size savings given this performance is possible with a 4-cavity filter.

Figure 5. Cavity filter requirements.

Figure 5. Cavity filter requirements.

Conclusion

In summary, both zero-IF and RF sampling architectures provide exceptional capability. However, where the goal is optimized cost, weight, and size, the zero-IF architecture wins on multiple accounts. From the perspective of power, the zero-IF architecture with integration of significant portions of analog gain offers a compelling power savings. Similarly, when considering the impact of filtering, zero-IF offers the potential to significantly downsize the filter requirements. While the cost differential of the filters may be small, the size and weight reduction of these filters should move beyond 50% based on the number of cavities required.

References

1 3GPP, 38.104 Rel 16, V16.5.0, 2020-09.

2 Ashkan Mashhour, William Domino, and Norman Beamish. “On the Direct Conversion Receiver—A Tutorial.” Microwave Journal, June 2001.

3 Jonathan Harris. “The ABCs of Interleaved ADCs.” Analog Devices, Inc., October 2019.

4 Brad Brannon. “AN-756: Sampled Systems and the Effects of Clock Phase Noise and Jitter.” Analog Devices, Inc., 2004.

5 Walt Kester. “MT-002: What the Nyquist Criterion Means to Your Sampled Data System Design.” Analog Devices, Inc., 2009.

6 Brad Brannon, Kenny Man, Nikhil Menon, and Ankit Gupta. “AN-1354: Integrated ZIF, RF to Bits, LTE, Wide Area Receiver Analysis and Test Results.” Analog Devices, Inc., July 2016.

7 Brad Brannon. “AN-410: Overcoming Converter Nonlinearities with Dither.” Analog Devices, Inc.

Author

Brad Brannon

Brad Brannon

Brad Brannon has worked at ADI for 37 years following his graduation from North Carolina State University. At ADI he has held positions in design, test, applications, and system engineering. Currently Brad is developing reference designs for O-RAN and supporting those customers. Brad has authored several articles and application notes on topics that span clocking data converters, designing radios, and testing ADCs.



from Hacker News https://ift.tt/MtmXrs3

No comments:

Post a Comment

Note: Only a member of this blog may post a comment.