For the past few years, many wireless designers have focused their efforts on developing direct-conversion receivers that reduce the bill-of-material (BOM) costs in next-generation radio architectures. The fruits of these efforts have been seen in the past year or so with a host of chip developers launching direct-conversion receivers to market.
Now it's time to bring this same level of integration to the transmit chain. By reworking their signal processing and converter architectures, designers can deliver direct-conversion transmitter solution that bring BOM costs to new levels in the radio front end.
Low IF vs. Zero IF
There are really two types of direct-conversion transmit architectures. The first is a zero IF upconversion where baseband signals are converted directly to RF signals. The second is a low-IF version where signals are synthesized at a low IF and then converted to an RF signal for transmission.
Figure 1 describes the frequency transformation that occurs in a zero-IF architecture. In this architecture, the modulating signal is centered around DC. When this signal is mixed with an LO, the result is a single spectrum centered around the LO frequency. As we will see later, any imperfections in the transmitter such as I/Q level imbalance, LO leakage, I/Q offsets and imperfect quadrature in the I/Q modulator will manifest themselves as imperfections in the desired RF spectrum and not as unwanted out-of-channel components (such as sidebands).
Click here for Figure 1
Figure 1: Diagram of a zero-IF baseband-to-RF direct conversion architecture.
Figure 2 highlights the low-IF architecture. Similar components are used in this architecture. However there is a key difference. The modulating signal from the DAC is no longer centered around dc but has been synthesized as an image-free low IF. When this signal is mixed with the LO in the I/Q modulator, a more complex spectrum results. The spectrum of the desired signal appears at an offset from the LO that is equal to the low IF frequency (i.e. (ωLO-ωIF).
Click here for Figure 2
Figure 2: Diagram of a low-IF direct-conversion transmit architecture.
As we will see later, imperfections in the I/Q modulator result in an unwanted LO leakage component along with an un-wanted upper sideband. However, by increasing the low IF frequency, these unwanted components can be moved far away enough from the desired spectrum to make filtering practical.
It should also be noted that modern DACs allow for active nulling of gain and offset mismatches. Calibrating the gain and offset mismatches reduces the unwanted out-of-band components (LO leakage and sideband image), making the filtering process that much easier. Assuming that these unwanted components can be adequately filtered, the net result of this scheme (compared to a baseband to RF conversion) is a perfectly modulated carrier.
While this second technique is, strictly speaking, not direct conversion (there is now an IF, albeit a very low one), because the architecture and component count is so similar to a zero-IF scheme and has similar economic benefits, we will still consider it a direct conversion architecture for the purposes of this paper.
Super-Nyquist Images
In direct-conversion architectures, reconstruction filters are required at the output of the DAC to reduce the amplitude of super-Nyquist images that occur as a result of the D/A conversion. Super-Nyquist images are images that appear as copies of the DAC spectral output between -Nyquist to +Nyquist at multiple intervals of the sample rate.
An example of a reconstruction DAC output spectrum is shown in Figure 3. The super-Nyquist images are attenuated by Sinc(a*f) function (where f is frequency).
Click here for Figure 3
Figure 3: Super-Nyquist images relative to the reconstruction DAC rate.
Since the super-Nyquist images are related to the sampling rate of the DAC, increasing the sampling rate of the DAC shifts these images further from the desired signal. This increases the transition band of the required reconstruction filter, thereby reducing its complexity, cost, and size. Increasing the transition band, however, creates a few issues for designs. These include:
- Increasing data rates from the digital backend to the DAC increases the drive requirements of the interface.
- It can result in an increase of high-frequency digital noise in the system.
- The digital backend power would increase.
- A new backend IC design could be required in a currently developed system.
One method to maintain the benefit, but address the issues above is to use a DAC with built-in interpolation filters. Interpolation filters will effectively increase the sample rate of the input signal while suppressing the original images sometimes by a factor of 70 dB or more.
The interpolation rate factor, N, determines how fast the actual DAC update rate (Fdac) will be relative to the input data rate (Fdata). For a system with 4x interpolation, Fdac equals 4 times Fdata, thereby shifting the closest super-Nyquist image to (N-1)*Fdata
Negative Images
Consider a single DAC being used to reconstruct a digitally modulated signal at some "low" IF, IFLOW, which is then upconverted to a higher IF (or RF) frequency using an external mixer with a input clock running at some frequency, LO. Since a single DAC can only output a "real" signal, the output signal will consist of the target signal somewhere between DC and Fdata/2 and a redundant negative image, a negative copy of the target signal between DC and --Fdata/2. This becomes an issue when the signal is upconverted and both spectral copies appear, centered on the LO frequency as shown in Figure 4a.
Click here for Figure 4
Figure 4: Representation of a.) Upconverted double sideband (DSB) signal and b.) upconverted single sideband (SSB) signal.
Since both the target signal and the image are redundant (i.e. contain the same information), to maximize spectral efficiency while reducing the transmit amplifier output requirements, either signal can be removed with no loss of information. To suppress the negative image by greater than 55 dB, filtering would be required at IF (using lossy, cascaded filters) or RF (using expensive, wideband cascaded filters) filter.
Some filtering will always be required at some point, but by using a single sideband (SSB) [also know as image rejection] architecture with a dual DAC, the negative image can be suppressed to reduce the requirements of the IF or RF filter. The SSB architecture uses complex data that represents a non-symmetrical signal spectrum. When the signal is upconverted with a quadrature modulator, the desired signal is present with the negative image attenuated, as shown in Figure 4b.
Depending on quadratue phase relationship (i.e. lead vs. lag) between the transmit DAC's output signals, the lower or higher image can be selected. The amount of image suppression realized at the low IF depends on various error sources. Here it is also helpful to have the ability to correct for gain mismatches between the quadrature signals. Compared to a DSB architecture, only a single IF (i.e. SAW) filter would be required to eliminate the residual image thus both reducing cost and board space.
Analog Quadrature Modulators
Figure 5 shows a block diagram of an ideal I/Q modulator. It is comprised of two mixers, a summer, and a phase splitter (or quadrature generator) which splits the LO signal into two components at the same frequency with a phase difference of exactly 90 degrees. To understand the operation of this component and its imperfections, let's consider what happens when we use it to perform SSB modulation, that is, low IF-to-RF conversion.
Click here for Figure 5
Figure 5: Diagram of a typical I/Q modulator.
Start with a low frequency modulating signal which we apply to the I and Q inputs as SinωBBt and CosωBBt. This signal is mixed with the quadrature components of a local oscillator (SinωLOt and CosωLOt) at 1900 MHz. Mathematically, the transformation can be represented as follows.
Vout = Cos(ωBBt) x Cos(ωLOt) + Sin(ωBBt) x Sin(ωLOt)
Using trigonometric identities, this equation simplifies to:
Vout = Sin((ωBB-ωLO)t)
If ωBB is a low IF, say 1 MHz and ωLO is at RF, say 1901 MHz, the resulting output signal is a SSB, suppressed carrier, signal at 1900 MHz as shown in Figure 6. (Note: if the baseband signal is centered at dc, the output signal will be centered at the LO frequency)
Click here for Figure 6
Figure 6. SSB modulation using an I/Q modulator.
LO Leakage
Now, let us assume that there is a slight voltage offset, VOSBB in one of the baseband (I/Q) input circuits (or in the drive signal from the DAC). From the original equation describing the transformation, we can now rewrite the equation for the baseband to RF transformation.
Vout = (Sin(ωBBt)+VOSBB) x Sin(ωLOt) + Cos(ωBBt) x Cos(ωLOt)
Simplifying this equation yields:
Vout = Sin((ωBB-ωLO)t) + VOSBB x Sin(ωLOt)
The result is the same as before except that now, there is a component of the output signal at the LO frequency, namely VOSBB x Sin(ωLOt). If the baseband signal is centered at dc, this component will fall in the center of the desired spectrum. So if this unwanted component is either at or very close to the desired output signal, RF filtering is not possible.
In practice, active offset nulling techniques are used to reduce this LO leakage. This necessitates a dc-coupled connection between the DAC and modulator. The offset nulling can occur in the digital backend, or at the output of the DAC. If done in the digital backend, an additional data processing step is required (not convenient in currently developed backends) and will directly reduce dynamic range of the data unless the resolution is increased. Alternatively, adding a small amount of CD offset to the DAC output signal, prior to the quadrature modulator, will compensate for channel offsets and reduce the LO leakage without reducing DAC dynamic range.
Using active nulling of offsets on the I and Q inputs at ambient, LO leakage can typically be held below about -50 dBm over temperature on a modulator which generates a maximum output power of around 0 dBm.
Because of the required dc-coupled link between DAC and modulator, the DAC/modulator interface should be designed so that bias levels and peak-to-peak voltages swings are compatible to avoid the need for expensive intermediate active stages.
It should be noted that the magnitude of the LO leakage is not related to the size of the modulator's output signal. Because the LO is internally limited before being applied to the mixers, the modulator's output amplitude is only proportional to the baseband drive levels. So if the baseband drive level needs to be reduced to accommodate, for example, a baseband signal with a high peak to average ratio, the modulator's output power will decrease proportionately. However the LO feedthrough level will stay the same, increasing the relative leakage (in dBc).
It's also important to note that since the LO leakage results from dc offset errors, nulling of LO leakage is independent of frequency to a first approximation. However as frequency increases, LO leakage to the output that results from other internal parasitic circuit elements, increases. Offset compensation will still reduce the overall feedthrough but the nulling will now become more frequency dependent.
Handling I/Q Imbalance
In the same way that offset errors in the baseband circuits led to increased LO leakage, baseband I/Q amplitude imbalance and imperfect quadrature and amplitude imbalances at the outputs of the phase splitter, create an unwanted upper sideband interferers (see Figure 6).
For example, a 0.2 dB amplitude imbalance and 1-deg. phase imbalance at the phase splitter outputs, results in upper sideband amplitude of -36 dBc (note that the size of this component is proportional to the output power of the desired signal). In general, there is no possibility to adjust the quadrature accuracy of the phase splitter. As a result, any trimming effort must focus on adjusting the relative amplitude and phase of the I and Q baseband inputs.
The Effect of Imperfections
Up to now we have discussed modulator imperfections in the context of SSB modulation for a lo-IF system. Under this approach, un-wanted products appear at some offset from the desired signal.
In zero IF architectures, on the other hand, the results of imperfection are embedded in the desired signal spectrum. Figures 7a and 7b, for example, shows the effect of excessive LO leakage and I/Q gain imbalance respectively on a 16-level quadrature amplitude modulated (QAM) signal that has been modulated from baseband (2 MSamples/s) up to an RF frequency of 900 MHz.
Click here for Figure 7a
Figure 7a: Effect of excessive LO leakage on the constellation of an upconverted QPSK signal.
Click here for Figure 7b
Figure 7b: Effect of I/Q drive imbalance on the constellation of an upconverted QPSK signal.
In the case of the excessive LO leakage, the constellation appears to become more circular. I/Q imbalance manifests itself, as one would expect, as a rectangular constellation. Clearly, there is no possibility to use filtering to improve signal quality. So to improve signal quality, we must rely on improved baseline component performance and/or active nulling of LO leakage and I/Q quadrature imbalances.
Dealing with Noise
One of the big challenges facing developers using direct-conversion radios is that the absence of IF frequencies in severely limits the capacity to perform narrowband filtering to remove noise and spurious components. Consider the case of a W-CDMA transmitter generating a single carrier (Figure 8).
Click here for Figure 8
Figure 8: W-CDMA output spectrum from a Direct Conversion transmitter.
In a frequency-agile system, the transmitter must be able to generate a 5 MHz channel anywhere within a 60 MHz bandwidth from 2110 MHz to 2170 MHz (channel raster for W-CDMA = 200 KHz). The only filtering that is permissible is narrowband lowpass filtering at baseband and band-filtering at RF (at the power amplifier output and possibly also at the modulator output).
Figure 8 shows how the resulting spectrum would look. Out-of-band noise and spurious components are effectively attenuated by the band filter. However, in-band we must live with any noise or distortion that the signal chain produces.
The adjacent channel power ratio (ACPR) and alternate channel power ratio are the measure of the noise and spurious components that exist in channels close the carrier. The level of ACPR and adjacent channel power ratio is related to the third-order intercept (IP3) of the signal chain and will become excessive as components in the signal chain are over-driven.
As we move away from the carrier, the spectrum becomes dominated by the noise floor of the signal chain. This noise floor is typically specified in terms of dBm/Hz (to calculate the in-band noise, add 10log(bandwidth) to specification) and is usually less than --150 dBm/Hz on good quality I/Q modulators.
From Figure 8, we can see that the key to optimizing signal-to-noise ratio (SNR) is to keep the in-band noise floor as low as possible while maximizing carrier power. However, increasing signal power will at some point begin to adversely affect ACPR so a balance must be struck. In addition, in a multi-carrier system, channel power will drop by 3 dB for each doubling of the number of carriers. Because the noise floor is fixed, adding carriers will therefore decrease SNR.
It should be noted that these in-band unwanted signal components (i.e. noise floor and ACPR) are equally prevalent and problematic in both zero-IF and low-IF to RF architectures.
Generating a W-CDMA carrier
Using the information above, let's generate a wideband CDMA (W-CDMA) carrier. Figure 9 shows the output spectrum of a direct-conversion modulator, driven from a 16-bit DAC. While the DAC also contains a complex modulator, allowing for the generation of an image-free low IF, in this case, it is generating a wideband-CDMA (W-CDMA) baseband spectrum centered at dc. This spectrum is mixed up to 1.85 GHz by driving the LO of the modulator with this frequency.
Click here for Figure 9
Figure 9: Spectrum of a W-CDMA signal up-converted from baseband to 1.85 GHz using a dual DAC and a direct-conversion modulator.
While the direct-conversion modulator has an output 1-dB compression of over +3dB, it is necessary to run the modulator output backed off to prevent clipping of the large peaks that are characteristic of a downlink (i.e. signal transmit from the basestation) W-CDMA signal. With a peak-to-average ratio (PAR) of approximately 14 dB, it is therefore necessary to run the output this far below the modulator's nominal output power to avoid excessive ACPR. The output power is regulated by varying the amplitude of the I/Q drive levels.
In the example detailed, the output channel power is -15dBm and the measured ACPR levels in the upper and lower channels are approximately -64 dB. Note that adjacent channel power, which is measured two channels away from the carrier, is much lower at approximately --68 dBc. Essentially, third-order intermodulation products dominate ACPR while the noise floor of the modulator dominates alternate channel power ratio. This can clearly be seen in Figure 9. In the adjacent channel, there is distinctive spectral re-growth but in the alternate channels, the spectrum is flat.
Wrap Up
With the drive on to continually cut size and cost, designers will continue to implement direct-conversion architectures in transmitters for W-CDMA and other emerging standards. But, developing these architectures will not be easy. Designers must first wrestle with tough technical challenges, such as LO leakage, to make direct-conversion transmitters and effective solution for wireless designs.
Editor's Note: This paper is based on a presentation made at the 2002 Communications Design Conference.
About the Author
Eamon Nash is an applications engineer at Analog Devices. He holds a Bachelor of Engineering degree (B.Eng) in Electronics from University of Limerick, Ireland. Eamon can be reached at eamon.nash@analog.com.
Anthony DeSimone is an applications engineer at Analog Devices. He holds a BSEE from University of Lowell, MA (1991) and an MSEE from Tufts University (1998). Anthony can be reached at anthony.desimone@analog.com.
