This article covers specific design choices in the creation of the world’s first K-band data converter, the EV12DS460A, explaining that by avoiding established CMOS design principles and applying an ultra-high speed process, new capabilities can emerge.

**Introduction**

Microwave system designers are constantly on the look out to improve performance and attain higher operating bandwidths. Design simplification and the reduction of power, size and weight are also desirable; UWB data converters offer significant system simplification for multi-channel transmit paths (for those unfamiliar with how, you can read our intro here). Many component developments have emerged over the last few years that help. However, few of these have the potential impact of a new digital-to-analog converter (DAC), the EV12DS460A. Indeed, the novel DAC claims to place complex bandwidth across a huge spectral range that includes the microwave K-band out to 26.5 GHz.

Initial details of a prototype monolithic microwave IC (MMIC) emerged during last year’s European Microwave IC conference. Early indications showed a product capable of turning in solid X-band performance (8 to 12 GHz). Subsequently, detailed wide-band testing revealed that this DAC is capable of lot more than that. The part operates across eight Nyquist zones with a low noise floor and minimal spurious signal generation.

The story here is of a device providing an early glimpse of a future where software defined microwave systems (SDeMS) become a reality. But in getting there, two obvious questions arise:

This article shows that by eschewing established CMOS design principles and applying an ultra-high speed process, that new capabilities emerge and how a compact, single-core quantizer coupled with careful design choices leads to breakthrough performance. You’ll discover how subtleties in layout and circuit simplifications are dominant design considerations. First up, let’s dip into the high-level architectural choices made.

**High level design**

Two factors determine the achievable performance namely:

Most high-speed DAC designs time interleave several cores to boost sample rate. However, their challenge comes in reconstructing the output signal. It’s hard to avoid spurious signal generation and the resultant dynamic degradation. The operation of interleaved DACs is not discussed here as their generally poor SFDR is the reason that an alternative segmented architecture was chosen for the described DAC.

**Segmented design**

Explaining basic DAC operation normally starts with the idea that a series of binary weighted current sources can be switched into a summing amplifier. Each ‘power of 2’ element is enabled or not, dependent on its relative bit position. The beauty of this implementation is simplicity and the limited number of elements (1 per bit) needed. In practice, its nigh impossible to scale sources linearly much beyond 8-bits.

Architecturally, there is one further simplification to help the quantizer core design. By adopting a hybrid segmented arrangement illustrated (fig 1). The segmented DAC divides the conversion task between an m-bit thermometer coding section and a second (n-m)-bit binary weighted section, handling fine (LSB) resolution. A time delay accounts for the thermometer encoding process after which the outputs from these two segments are summed for the final multi-bit result.

As noted previously, attaining better than 8-bit linearity is a hard-won battle, but by dividing a multiple bit convertor up into most significant (MSB) and least significant bit (LSB) segments minimizes matching and lowers core complexity. With careful design, it is possible to construct both the thermometer and binary weighted segments from identical switches, resistors and current sources.

**Single core simplicity**

Delivering good static accuracy is the starting point for any convertor design, in the hybrid-segmented approach, precision is defined by the tolerances achievable in the binary weighted LSB section.

A design goal was to deliver high SFDR and avoid calibration, mandating the achievement of better than 0.5 LSB matching. Three potential quantizer configurations were considered:

- 2-bit thermometer (= 3 segments) with 10-bit weighted segments = 13 segments
- 3-bit thermometer (= 7 segments) with 9-bit weighted segments = 16 segments
- 4-bit thermometer (= 15 segments) with 8-bit weighted segments = 23 segments

Initial inspection suggests option 1 is the obvious design choice; it promises the lowest number of segments and hence the smallest core area. However, its static accuracy lets it down. To understand this, consider that a 12-bit quantizer able to output a full-scale 1 volt peak to peak implies a LSB quantization voltage of 244 μV (1Vp-p/4096). Simulation shows that the matching achievable with 9-bit weighted segments is 125 μV. That’s 2x better than this (i.e. 0.5 LSB at 12-bits) which guarantees monotonic DAC operation. However, with the ‘obvious’ 10-bit weighted option, it is not possible to improve matching, 125 μV is the physical process limit, thus option 1 is unviable. Simulation also eliminated the third option as it is compromised by the excessive dynamic loading it places on the clock buffer.

**Process technology**

Designing the DAC eschewed standard CMOS processes in favor of taking the path less trodden. This philosophy exploits controlling the raw speed of a heterojunction silicon-germanium carbon (SiGeC) bipolar process sourced from Infineon^{1}. By introducing carbon inside the intrinsic base of the NPN bipolar transistors, the B7HF200 process allows for a thin and highly doped base. It is a key element to achieve a high transition frequency (Ft of 200 GHz) and a low base resistance, two parameters on which the DAC’s performance strongly relies.

This process has served high speed and mm-wave applications well for nearly a decade now and is applied in several solid-state microwave components.

B7HF200’s speed is further enhanced by the provision of four layers of copper metal suitable for low current density interconnects. Copper helps minimize undesirable circuit parasitics, the bane of high speed designers.

**DAC design secrets**

Designers did not achieve the blistering speed of the EV12DS460A in a single serendipitous leap. Its architecture evolved over several generations since the introduction of a slower 12-bitter in 2011^{2}. Even that earlier part demonstrated world-class performance; generating a 1.5 GHz bandwidth.

The design journey is focused on three general design principles:

**Driving the quantizer’s dynamic load**

The quantizer’s design is partially reproduced (fig 3). On the right is the quantizer comprising 16 segments, whilst on the left, is the analog circuitry of the sample clock system. Lumped together, bridging the two halves of the circuit are parasitics that arise from on chip traces and represented here by Lp and Cp.

To support sampling at 6 to 7 GSps, it is important to have a low jitter clock source with ultra-fast transitions. At 6 GSps the clock period is only 166 ps. Ensuring clean, rapid transitions is paramount to enable fast quantizer settling and thus sampling. However, the relatively high, full-scale quantizer current is set at 20 mA in this design. To drive this quickly, demands a sophisticated driver comprising the differential pair and output follower circuit which features extremely low output impedance.

For such a driver circuit, the output impedance Zout can be expressed as:

Zout = (1/g_{m} + R_{bb} + R_{g})/Beta(f), where g_{m} is the transistor transconductance (1/g_{m}=1,25 ohms), R_{bb} is the follower output resistance, R_{g} the output resistance of the differential pair and Beta(f) is the dynamic current gain of the transistors versus frequency.

Considering the B7HF200 process characteristics (cutoff frequency f_{T} = 200 GHz), the current gain Beta(f) at 20 GHz is equal to 10. Also, the very low bipolar transistor intrinsic base resistance delivers an R_{bb} in triple base configuration of 25 ohms.

R_{g} should also be minimized as much as practical with the constraint of maintaining a sufficient value to avoid increasing the bias current too much and consequently the power consumption. A value of roughly 50 ohms was obtained.

Finally, a first order estimate for driver output impedance is: Zout = (1.25 + 25 + 50)/10 = ~ 7.5 ohms. This low output impedance is a key element for fast operation.

Maintaining a 300 mV pulse amplitude at the output buffer demands driving 300 mV across 50 ohm termination (300 mV/50 = 6 mA). Further improvements in R_{g} make only modest impedance improvements at a cost of higher power consumption. Halving Rg, the bias current must rise to 12 mA.

**Minimized trace lengths and guaranteed DAC stability**

Back to the importance of trace length and its impact on parasitics in high-speed design; each quantizer segment of the described design is only 50 μm wide, so its modest sixteen segments combine to form a total signal trace length of 800 μm (16 x 50 μm), any reduction in pitch is therefore helpful.

The EV12DS460A’s global time constant can be factored from three contributors:

- The dynamic load capacitance (C
_{L}) estimated at 0.5 pF (C_{L}=g_{m.}T_{f}with g_{m}= ΔI/ΔV = ~ 20mA/25mV and T_{f}the transistor forward transit time = 0,8 ps) - The passive parasitic capacitance(C
_{P}) of the metal signal trace estimated at 0.5 pF - The passive parasitic inductance (L
_{P}) of the metal trace estimated at 50 pH

Under worst case conditions, the global time constant ΣT can be calculated as follows:

ΣT=Zout.C_{L} +Zout.C_{P} +L_{P}/Zout,soΣT=7.5Ω.0.5pF+7.5Ω.500fF+50pH/7.5Ω=3.75ps+3.75ps+6.66ps
= **~14 ps**

This time constant correlates well with measured DAC data of 35 ps rise and fall times (tr/tf). Furthermore, at this level, tr/tf individually represent less than 20% of the total sample clock period (of 166 ps) giving fast enough clock edges to support a first order bandwidth approximation of 10 GHz, hitting the DAC’s design goal.

Beyond this first order assessment, some special damping techniques were used to ensure dynamic stability within the DAC. Maximum overshoot (+4%) and minimal ring back (-2%) were achieved. Certainly, the fact that the B7HF200 process offers low sheet resistance, copper metallization helps further tweak and damp critical chip nodes. The performance of the resulting exceptionally clean, 6 GHz sampling is illustrated in the step response (fig 4).

**Dynamic enhancement through output pulse shaping**

Four output pulse shaping modes (NRZ, NRTZ, RTZ, RF) are provided to give system designers freedom to tailor the DAC’s dynamic response to specific output frequency bands, thus facilitating frequency planning. Most quantizer distortion can be tracked down to switching transitions. Any switching glitches are ultimately superimposed on the output signal (fig 5), however if these glitches can be removed, then the output spectral purity benefits.

This is achieved in the pulse shaped trace above, by forcing a return to zero ahead of each edge transition, visible here for both the NRTZ and RF modes. Pulse shaping is programmed via a 3-wire serial interface with two user shaping controls, the re-shaping pulse width (RPW) and pulse centering (RPB). The pulse center must coincide with the center of the transition edge if all glitch energy is to be removed. Note that this technique sacrifices a small amount of output signal power (equivalent to the area defined by RPW).

Characteristic curves (fig 6) illustrate the benefits of pulse shaping. This data shows the frequency spectrum across eight Nyquist zones out to 27 GHz (for both fs = 6 & 7 GSps) at two RPW settings (for those unfamiliar with signal aliasing read our overview here). Note that increasing sample rate noticeably expands the typical SINC (sin(x)/x) DAC output characteristic.

Up to + 12 dB improvement in harmonics levels is shown for the third harmonic due to reshaping (H3 improved from -57 dBm to -69 dBm), stretching the DAC’s ‘reach’. To further emphasize this, the following spectra (fig 7) has been produced at 6 GSps with Fout = 2940 MHz, both with (NRTZ mode) and without reshaping (NRZ mode). In NRTZ mode, the benefit of re-shaping is clearly visible.

**Measured Performance**

Output 3 dB bandwidth extends up to 7 GHz, with a guaranteed sample rate of 6 GSps enabling the generation of a 3 GHz wide, complex bandwidth. Usable output power is clearly visible in the X-band (fig 8a). The curve shows a single tone carrier at 11950 MHz with a SFDR of 50 dBc in the 4th Nyquist zone. Here the 4th harmonic dominates SFDR. This carrier frequency was carefully selected to lie on the edge of the X-band allowing many of the harmonic signals to be easily spotted as they occur in their natural harmonic order.

Increasing carrier frequency into the K-band (fig 8), with a signal reference now set at 23950 MHz in the 8th Nyquist zone, the 2nd harmonic dominates SFDR (-36.5 dBc). Again, a clean progression of harmonics is visible.

Other performance characteristics standout from these two curves, both plots show non-harmonic spurs at the mid-frequency point. These spurs can be traced back to insufficient crossover signal rejection within the DAC’s 4:1 input multiplexer. Even so, the spurs peak at -80 dBm, a very respectable level. The DAC’s noise floor is measured at close to -110 dBm.

Testing data convertor products with single or multi-tones is easily arranged in the lab. Judging the capability of a DAC on the strength of these tests, only partly tells the story. Today’s digital communication systems deploy complex modulation on large chunks of bandwidth, and so a more representative broadband test is desirable. That’s where noise o power ratio (NPR) testing helps; it exercises a DAC across a wide bandwidth and can gauge how a signal comprising of many incoherent and narrow band tones, interact and interfere with one another when mixed by the DAC. A DAC with an NPR closely matching the ideal NPR of an n-bit device is clearly a desirable broadband component.

Broadband NPR for this part is shown below (fig 9). Sampling at 7 GSps allows the generation of 3.150 GHz wide synthetic bandwidth. The resultant NPR is 42.6 dB, equivalent to an 8.6 effective number of bits (ENOB). Note the considerable NPR flatness out to 3325 MHz. The NPR test is usually performed using a digital pattern with a Gaussian noise power density. A (digital) notch filter is applied to the pattern in the frequency domain to give a ‘quiet’ zone within the bandwidth of interest. The pattern is sent to the DAC and the NPR performance is calculated as the ratio of average magnitudes of power densities measured both within and outside the notch. For an ideal DAC, the signal power within the notch is related to quantization noise only. For a real DAC, the quantization noise is combined with DAC thermal and voltage noise induced by clock jitter, (output referred) plus intermodulation products due to any cross-channel interferers.

The second NPR characteristic (fig 10) replicates a 3.150/2.700 GHz NPR pattern across + 22 GHz using 7/6 GSps sampling with the DAC operating in RF mode. This graph pair helps to highlight one of the benefits of increased sample rate. Not only does it impact the maximum bandwidth generation capability of a DAC, but it expands the SINC characteristic and the output power available in high order Nyquist zones.

**Other state-of-the-art DACs**

Texas Instruments recently described a 14-bit 8.9 GSps RF DAC using 40 nm CMOS process supporting 4G LTE applications. It features an SFDR of 50 dBc at 8.9 GSps (Fout = 4300 MHz)^{3}. Although the DAC is capable of sampling at 8.9 GSps, no measurement results are available for frequencies above 4300 MHz excluding it from most microwave bands.

Another development is Analog Devices’ 11/16-Bit, 12 GSps DACs (the AD9161/AD9162). The AD9161/62 specifies sampling rate at 12 GSps in RF mode (also called mixed mode). In RF mode, since data is inverted every half a clock period, it looks like the DAC is sampling at 12 GSps. For the EV12DS460A operating in RF (see Fig. 5), the Data inversion is not considered in the specified sampling rate (6 GSps). Therefore, the EV12DS460A and AD9161/62 are strictly equivalent in terms of sampling rate. This is validated by the fact that for all, the instantaneous bandwidth is 3 GHz.

Both Analog Devices parts feature excellent SFDR across the first two Nyquist zones of 65 dBc (Fclock = 5 GSps, Fout = 4000 MHz). Unfortunately, performance collapses at frequencies above 7500 MHz. Output power rolls off to -66 dBm (Fout = 7500 MHz) preventing them doing useful work in X and K-bands.

**Final thoughts**

With the arrival of the EV12DS460A, microwave engineers now have a practical broad-band DAC capable of projecting complex bandwidth from DC all the way into the K-band frequency. Certainly, the device is not the only giga-sample DAC available, but as shown here, it is the first one capable of projecting large synthetic bandwidth into higher Nyquist zones whilst maintaining good spectral purity. It opens a realm of exciting and innovative possibilities for new mm-wave applications.

**Introduction**

Microwave system designers are constantly on the look out to improve performance and attain higher operating bandwidths. Design simplification and the reduction of power, size and weight are also desirable; UWB data converters offer significant system simplification for multi-channel transmit paths (for those unfamiliar with how, you can read our intro here). Many component developments have emerged over the last few years that help. However, few of these have the potential impact of a new digital-to-analog converter (DAC), the EV12DS460A. Indeed, the novel DAC claims to place complex bandwidth across a huge spectral range that includes the microwave K-band out to 26.5 GHz.

Initial details of a prototype monolithic microwave IC (MMIC) emerged during last year’s European Microwave IC conference. Early indications showed a product capable of turning in solid X-band performance (8 to 12 GHz). Subsequently, detailed wide-band testing revealed that this DAC is capable of lot more than that. The part operates across eight Nyquist zones with a low noise floor and minimal spurious signal generation.

The story here is of a device providing an early glimpse of a future where software defined microwave systems (SDeMS) become a reality. But in getting there, two obvious questions arise:

This article shows that by eschewing established CMOS design principles and applying an ultra-high speed process, that new capabilities emerge and how a compact, single-core quantizer coupled with careful design choices leads to breakthrough performance. You’ll discover how subtleties in layout and circuit simplifications are dominant design considerations. First up, let’s dip into the high-level architectural choices made.

**High level design**

Two factors determine the achievable performance namely:

Most high-speed DAC designs time interleave several cores to boost sample rate. However, their challenge comes in reconstructing the output signal. It’s hard to avoid spurious signal generation and the resultant dynamic degradation. The operation of interleaved DACs is not discussed here as their generally poor SFDR is the reason that an alternative segmented architecture was chosen for the described DAC.

**Segmented design**

Explaining basic DAC operation normally starts with the idea that a series of binary weighted current sources can be switched into a summing amplifier. Each ‘power of 2’ element is enabled or not, dependent on its relative bit position. The beauty of this implementation is simplicity and the limited number of elements (1 per bit) needed. In practice, its nigh impossible to scale sources linearly much beyond 8-bits.

Architecturally, there is one further simplification to help the quantizer core design. By adopting a hybrid segmented arrangement illustrated (fig 1). The segmented DAC divides the conversion task between an m-bit thermometer coding section and a second (n-m)-bit binary weighted section, handling fine (LSB) resolution. A time delay accounts for the thermometer encoding process after which the outputs from these two segments are summed for the final multi-bit result.

As noted previously, attaining better than 8-bit linearity is a hard-won battle, but by dividing a multiple bit convertor up into most significant (MSB) and least significant bit (LSB) segments minimizes matching and lowers core complexity. With careful design, it is possible to construct both the thermometer and binary weighted segments from identical switches, resistors and current sources.

**Single core simplicity**

Delivering good static accuracy is the starting point for any convertor design, in the hybrid-segmented approach, precision is defined by the tolerances achievable in the binary weighted LSB section.

A design goal was to deliver high SFDR and avoid calibration, mandating the achievement of better than 0.5 LSB matching. Three potential quantizer configurations were considered:

- 2-bit thermometer (= 3 segments) with 10-bit weighted segments = 13 segments
- 3-bit thermometer (= 7 segments) with 9-bit weighted segments = 16 segments
- 4-bit thermometer (= 15 segments) with 8-bit weighted segments = 23 segments

Initial inspection suggests option 1 is the obvious design choice; it promises the lowest number of segments and hence the smallest core area. However, its static accuracy lets it down. To understand this, consider that a 12-bit quantizer able to output a full-scale 1 volt peak to peak implies a LSB quantization voltage of 244 μV (1Vp-p/4096). Simulation shows that the matching achievable with 9-bit weighted segments is 125 μV. That’s 2x better than this (i.e. 0.5 LSB at 12-bits) which guarantees monotonic DAC operation. However, with the ‘obvious’ 10-bit weighted option, it is not possible to improve matching, 125 μV is the physical process limit, thus option 1 is unviable. Simulation also eliminated the third option as it is compromised by the excessive dynamic loading it places on the clock buffer.

**Process technology**

Designing the DAC eschewed standard CMOS processes in favor of taking the path less trodden. This philosophy exploits controlling the raw speed of a heterojunction silicon-germanium carbon (SiGeC) bipolar process sourced from Infineon^{1}. By introducing carbon inside the intrinsic base of the NPN bipolar transistors, the B7HF200 process allows for a thin and highly doped base. It is a key element to achieve a high transition frequency (Ft of 200 GHz) and a low base resistance, two parameters on which the DAC’s performance strongly relies.

This process has served high speed and mm-wave applications well for nearly a decade now and is applied in several solid-state microwave components.

B7HF200’s speed is further enhanced by the provision of four layers of copper metal suitable for low current density interconnects. Copper helps minimize undesirable circuit parasitics, the bane of high speed designers.

**DAC design secrets**

Designers did not achieve the blistering speed of the EV12DS460A in a single serendipitous leap. Its architecture evolved over several generations since the introduction of a slower 12-bitter in 2011^{2}. Even that earlier part demonstrated world-class performance; generating a 1.5 GHz bandwidth.

The design journey is focused on three general design principles:

**Driving the quantizer’s dynamic load**

The quantizer’s design is partially reproduced (fig 3). On the right is the quantizer comprising 16 segments, whilst on the left, is the analog circuitry of the sample clock system. Lumped together, bridging the two halves of the circuit are parasitics that arise from on chip traces and represented here by Lp and Cp.

To support sampling at 6 to 7 GSps, it is important to have a low jitter clock source with ultra-fast transitions. At 6 GSps the clock period is only 166 ps. Ensuring clean, rapid transitions is paramount to enable fast quantizer settling and thus sampling. However, the relatively high, full-scale quantizer current is set at 20 mA in this design. To drive this quickly, demands a sophisticated driver comprising the differential pair and output follower circuit which features extremely low output impedance.

For such a driver circuit, the output impedance Zout can be expressed as:

Zout = (1/g_{m} + R_{bb} + R_{g})/Beta(f), where g_{m} is the transistor transconductance (1/g_{m}=1,25 ohms), R_{bb} is the follower output resistance, R_{g} the output resistance of the differential pair and Beta(f) is the dynamic current gain of the transistors versus frequency.

Considering the B7HF200 process characteristics (cutoff frequency f_{T} = 200 GHz), the current gain Beta(f) at 20 GHz is equal to 10. Also, the very low bipolar transistor intrinsic base resistance delivers an R_{bb} in triple base configuration of 25 ohms.

R_{g} should also be minimized as much as practical with the constraint of maintaining a sufficient value to avoid increasing the bias current too much and consequently the power consumption. A value of roughly 50 ohms was obtained.

Finally, a first order estimate for driver output impedance is: Zout = (1.25 + 25 + 50)/10 = ~ 7.5 ohms. This low output impedance is a key element for fast operation.

Maintaining a 300 mV pulse amplitude at the output buffer demands driving 300 mV across 50 ohm termination (300 mV/50 = 6 mA). Further improvements in R_{g} make only modest impedance improvements at a cost of higher power consumption. Halving Rg, the bias current must rise to 12 mA.

**Minimized trace lengths and guaranteed DAC stability**

Back to the importance of trace length and its impact on parasitics in high-speed design; each quantizer segment of the described design is only 50 μm wide, so its modest sixteen segments combine to form a total signal trace length of 800 μm (16 x 50 μm), any reduction in pitch is therefore helpful.

The EV12DS460A’s global time constant can be factored from three contributors:

- The dynamic load capacitance (C
_{L}) estimated at 0.5 pF (C_{L}=g_{m.}T_{f}with g_{m}= ΔI/ΔV = ~ 20mA/25mV and T_{f}the transistor forward transit time = 0,8 ps) - The passive parasitic capacitance(C
_{P}) of the metal signal trace estimated at 0.5 pF - The passive parasitic inductance (L
_{P}) of the metal trace estimated at 50 pH

Under worst case conditions, the global time constant ΣT can be calculated as follows:

ΣT=Zout.C_{L} +Zout.C_{P} +L_{P}/Zout,soΣT=7.5Ω.0.5pF+7.5Ω.500fF+50pH/7.5Ω=3.75ps+3.75ps+6.66ps
= **~14 ps**

This time constant correlates well with measured DAC data of 35 ps rise and fall times (tr/tf). Furthermore, at this level, tr/tf individually represent less than 20% of the total sample clock period (of 166 ps) giving fast enough clock edges to support a first order bandwidth approximation of 10 GHz, hitting the DAC’s design goal.

Beyond this first order assessment, some special damping techniques were used to ensure dynamic stability within the DAC. Maximum overshoot (+4%) and minimal ring back (-2%) were achieved. Certainly, the fact that the B7HF200 process offers low sheet resistance, copper metallization helps further tweak and damp critical chip nodes. The performance of the resulting exceptionally clean, 6 GHz sampling is illustrated in the step response (fig 4).

**Dynamic enhancement through output pulse shaping**

Four output pulse shaping modes (NRZ, NRTZ, RTZ, RF) are provided to give system designers freedom to tailor the DAC’s dynamic response to specific output frequency bands, thus facilitating frequency planning. Most quantizer distortion can be tracked down to switching transitions. Any switching glitches are ultimately superimposed on the output signal (fig 5), however if these glitches can be removed, then the output spectral purity benefits.

This is achieved in the pulse shaped trace above, by forcing a return to zero ahead of each edge transition, visible here for both the NRTZ and RF modes. Pulse shaping is programmed via a 3-wire serial interface with two user shaping controls, the re-shaping pulse width (RPW) and pulse centering (RPB). The pulse center must coincide with the center of the transition edge if all glitch energy is to be removed. Note that this technique sacrifices a small amount of output signal power (equivalent to the area defined by RPW).

Characteristic curves (fig 6) illustrate the benefits of pulse shaping. This data shows the frequency spectrum across eight Nyquist zones out to 27 GHz (for both fs = 6 & 7 GSps) at two RPW settings (for those unfamiliar with signal aliasing read our overview here). Note that increasing sample rate noticeably expands the typical SINC (sin(x)/x) DAC output characteristic.

Up to + 12 dB improvement in harmonics levels is shown for the third harmonic due to reshaping (H3 improved from -57 dBm to -69 dBm), stretching the DAC’s ‘reach’. To further emphasize this, the following spectra (fig 7) has been produced at 6 GSps with Fout = 2940 MHz, both with (NRTZ mode) and without reshaping (NRZ mode). In NRTZ mode, the benefit of re-shaping is clearly visible.

**Measured Performance**

Output 3 dB bandwidth extends up to 7 GHz, with a guaranteed sample rate of 6 GSps enabling the generation of a 3 GHz wide, complex bandwidth. Usable output power is clearly visible in the X-band (fig 8a). The curve shows a single tone carrier at 11950 MHz with a SFDR of 50 dBc in the 4th Nyquist zone. Here the 4th harmonic dominates SFDR. This carrier frequency was carefully selected to lie on the edge of the X-band allowing many of the harmonic signals to be easily spotted as they occur in their natural harmonic order.

Increasing carrier frequency into the K-band (fig 8), with a signal reference now set at 23950 MHz in the 8th Nyquist zone, the 2nd harmonic dominates SFDR (-36.5 dBc). Again, a clean progression of harmonics is visible.

Other performance characteristics standout from these two curves, both plots show non-harmonic spurs at the mid-frequency point. These spurs can be traced back to insufficient crossover signal rejection within the DAC’s 4:1 input multiplexer. Even so, the spurs peak at -80 dBm, a very respectable level. The DAC’s noise floor is measured at close to -110 dBm.

Testing data convertor products with single or multi-tones is easily arranged in the lab. Judging the capability of a DAC on the strength of these tests, only partly tells the story. Today’s digital communication systems deploy complex modulation on large chunks of bandwidth, and so a more representative broadband test is desirable. That’s where noise o power ratio (NPR) testing helps; it exercises a DAC across a wide bandwidth and can gauge how a signal comprising of many incoherent and narrow band tones, interact and interfere with one another when mixed by the DAC. A DAC with an NPR closely matching the ideal NPR of an n-bit device is clearly a desirable broadband component.

Broadband NPR for this part is shown below (fig 9). Sampling at 7 GSps allows the generation of 3.150 GHz wide synthetic bandwidth. The resultant NPR is 42.6 dB, equivalent to an 8.6 effective number of bits (ENOB). Note the considerable NPR flatness out to 3325 MHz. The NPR test is usually performed using a digital pattern with a Gaussian noise power density. A (digital) notch filter is applied to the pattern in the frequency domain to give a ‘quiet’ zone within the bandwidth of interest. The pattern is sent to the DAC and the NPR performance is calculated as the ratio of average magnitudes of power densities measured both within and outside the notch. For an ideal DAC, the signal power within the notch is related to quantization noise only. For a real DAC, the quantization noise is combined with DAC thermal and voltage noise induced by clock jitter, (output referred) plus intermodulation products due to any cross-channel interferers.

The second NPR characteristic (fig 10) replicates a 3.150/2.700 GHz NPR pattern across + 22 GHz using 7/6 GSps sampling with the DAC operating in RF mode. This graph pair helps to highlight one of the benefits of increased sample rate. Not only does it impact the maximum bandwidth generation capability of a DAC, but it expands the SINC characteristic and the output power available in high order Nyquist zones.

**Other state-of-the-art DACs**

Texas Instruments recently described a 14-bit 8.9 GSps RF DAC using 40 nm CMOS process supporting 4G LTE applications. It features an SFDR of 50 dBc at 8.9 GSps (Fout = 4300 MHz)^{3}. Although the DAC is capable of sampling at 8.9 GSps, no measurement results are available for frequencies above 4300 MHz excluding it from most microwave bands.

Another development is Analog Devices’ 11/16-Bit, 12 GSps DACs (the AD9161/AD9162). The AD9161/62 specifies sampling rate at 12 GSps in RF mode (also called mixed mode). In RF mode, since data is inverted every half a clock period, it looks like the DAC is sampling at 12 GSps. For the EV12DS460A operating in RF (see Fig. 5), the Data inversion is not considered in the specified sampling rate (6 GSps). Therefore, the EV12DS460A and AD9161/62 are strictly equivalent in terms of sampling rate. This is validated by the fact that for all, the instantaneous bandwidth is 3 GHz.

Both Analog Devices parts feature excellent SFDR across the first two Nyquist zones of 65 dBc (Fclock = 5 GSps, Fout = 4000 MHz). Unfortunately, performance collapses at frequencies above 7500 MHz. Output power rolls off to -66 dBm (Fout = 7500 MHz) preventing them doing useful work in X and K-bands.

**Final thoughts**

With the arrival of the EV12DS460A, microwave engineers now have a practical broad-band DAC capable of projecting complex bandwidth from DC all the way into the K-band frequency. Certainly, the device is not the only giga-sample DAC available, but as shown here, it is the first one capable of projecting large synthetic bandwidth into higher Nyquist zones whilst maintaining good spectral purity. It opens a realm of exciting and innovative possibilities for new mm-wave applications.

**Introduction**

Microwave system designers are constantly on the look out to improve performance and attain higher operating bandwidths. Design simplification and the reduction of power, size and weight are also desirable; UWB data converters offer significant system simplification for multi-channel transmit paths (for those unfamiliar with how, you can read our intro here). Many component developments have emerged over the last few years that help. However, few of these have the potential impact of a new digital-to-analog converter (DAC), the EV12DS460A. Indeed, the novel DAC claims to place complex bandwidth across a huge spectral range that includes the microwave K-band out to 26.5 GHz.

Initial details of a prototype monolithic microwave IC (MMIC) emerged during last year’s European Microwave IC conference. Early indications showed a product capable of turning in solid X-band performance (8 to 12 GHz). Subsequently, detailed wide-band testing revealed that this DAC is capable of lot more than that. The part operates across eight Nyquist zones with a low noise floor and minimal spurious signal generation.

The story here is of a device providing an early glimpse of a future where software defined microwave systems (SDeMS) become a reality. But in getting there, two obvious questions arise:

This article shows that by eschewing established CMOS design principles and applying an ultra-high speed process, that new capabilities emerge and how a compact, single-core quantizer coupled with careful design choices leads to breakthrough performance. You’ll discover how subtleties in layout and circuit simplifications are dominant design considerations. First up, let’s dip into the high-level architectural choices made.

**High level design**

Two factors determine the achievable performance namely:

Most high-speed DAC designs time interleave several cores to boost sample rate. However, their challenge comes in reconstructing the output signal. It’s hard to avoid spurious signal generation and the resultant dynamic degradation. The operation of interleaved DACs is not discussed here as their generally poor SFDR is the reason that an alternative segmented architecture was chosen for the described DAC.

**Segmented design**

Explaining basic DAC operation normally starts with the idea that a series of binary weighted current sources can be switched into a summing amplifier. Each ‘power of 2’ element is enabled or not, dependent on its relative bit position. The beauty of this implementation is simplicity and the limited number of elements (1 per bit) needed. In practice, its nigh impossible to scale sources linearly much beyond 8-bits.

Architecturally, there is one further simplification to help the quantizer core design. By adopting a hybrid segmented arrangement illustrated (fig 1). The segmented DAC divides the conversion task between an m-bit thermometer coding section and a second (n-m)-bit binary weighted section, handling fine (LSB) resolution. A time delay accounts for the thermometer encoding process after which the outputs from these two segments are summed for the final multi-bit result.

As noted previously, attaining better than 8-bit linearity is a hard-won battle, but by dividing a multiple bit convertor up into most significant (MSB) and least significant bit (LSB) segments minimizes matching and lowers core complexity. With careful design, it is possible to construct both the thermometer and binary weighted segments from identical switches, resistors and current sources.

**Single core simplicity**

Delivering good static accuracy is the starting point for any convertor design, in the hybrid-segmented approach, precision is defined by the tolerances achievable in the binary weighted LSB section.

A design goal was to deliver high SFDR and avoid calibration, mandating the achievement of better than 0.5 LSB matching. Three potential quantizer configurations were considered:

- 2-bit thermometer (= 3 segments) with 10-bit weighted segments = 13 segments
- 3-bit thermometer (= 7 segments) with 9-bit weighted segments = 16 segments
- 4-bit thermometer (= 15 segments) with 8-bit weighted segments = 23 segments

Initial inspection suggests option 1 is the obvious design choice; it promises the lowest number of segments and hence the smallest core area. However, its static accuracy lets it down. To understand this, consider that a 12-bit quantizer able to output a full-scale 1 volt peak to peak implies a LSB quantization voltage of 244 μV (1Vp-p/4096). Simulation shows that the matching achievable with 9-bit weighted segments is 125 μV. That’s 2x better than this (i.e. 0.5 LSB at 12-bits) which guarantees monotonic DAC operation. However, with the ‘obvious’ 10-bit weighted option, it is not possible to improve matching, 125 μV is the physical process limit, thus option 1 is unviable. Simulation also eliminated the third option as it is compromised by the excessive dynamic loading it places on the clock buffer.

**Process technology**

Designing the DAC eschewed standard CMOS processes in favor of taking the path less trodden. This philosophy exploits controlling the raw speed of a heterojunction silicon-germanium carbon (SiGeC) bipolar process sourced from Infineon^{1}. By introducing carbon inside the intrinsic base of the NPN bipolar transistors, the B7HF200 process allows for a thin and highly doped base. It is a key element to achieve a high transition frequency (Ft of 200 GHz) and a low base resistance, two parameters on which the DAC’s performance strongly relies.

This process has served high speed and mm-wave applications well for nearly a decade now and is applied in several solid-state microwave components.

B7HF200’s speed is further enhanced by the provision of four layers of copper metal suitable for low current density interconnects. Copper helps minimize undesirable circuit parasitics, the bane of high speed designers.

**DAC design secrets**

Designers did not achieve the blistering speed of the EV12DS460A in a single serendipitous leap. Its architecture evolved over several generations since the introduction of a slower 12-bitter in 2011^{2}. Even that earlier part demonstrated world-class performance; generating a 1.5 GHz bandwidth.

The design journey is focused on three general design principles:

**Driving the quantizer’s dynamic load**

The quantizer’s design is partially reproduced (fig 3). On the right is the quantizer comprising 16 segments, whilst on the left, is the analog circuitry of the sample clock system. Lumped together, bridging the two halves of the circuit are parasitics that arise from on chip traces and represented here by Lp and Cp.

To support sampling at 6 to 7 GSps, it is important to have a low jitter clock source with ultra-fast transitions. At 6 GSps the clock period is only 166 ps. Ensuring clean, rapid transitions is paramount to enable fast quantizer settling and thus sampling. However, the relatively high, full-scale quantizer current is set at 20 mA in this design. To drive this quickly, demands a sophisticated driver comprising the differential pair and output follower circuit which features extremely low output impedance.

For such a driver circuit, the output impedance Zout can be expressed as:

Zout = (1/g_{m} + R_{bb} + R_{g})/Beta(f), where g_{m} is the transistor transconductance (1/g_{m}=1,25 ohms), R_{bb} is the follower output resistance, R_{g} the output resistance of the differential pair and Beta(f) is the dynamic current gain of the transistors versus frequency.

Considering the B7HF200 process characteristics (cutoff frequency f_{T} = 200 GHz), the current gain Beta(f) at 20 GHz is equal to 10. Also, the very low bipolar transistor intrinsic base resistance delivers an R_{bb} in triple base configuration of 25 ohms.

R_{g} should also be minimized as much as practical with the constraint of maintaining a sufficient value to avoid increasing the bias current too much and consequently the power consumption. A value of roughly 50 ohms was obtained.

Finally, a first order estimate for driver output impedance is: Zout = (1.25 + 25 + 50)/10 = ~ 7.5 ohms. This low output impedance is a key element for fast operation.

Maintaining a 300 mV pulse amplitude at the output buffer demands driving 300 mV across 50 ohm termination (300 mV/50 = 6 mA). Further improvements in R_{g} make only modest impedance improvements at a cost of higher power consumption. Halving Rg, the bias current must rise to 12 mA.

**Minimized trace lengths and guaranteed DAC stability**

Back to the importance of trace length and its impact on parasitics in high-speed design; each quantizer segment of the described design is only 50 μm wide, so its modest sixteen segments combine to form a total signal trace length of 800 μm (16 x 50 μm), any reduction in pitch is therefore helpful.

The EV12DS460A’s global time constant can be factored from three contributors:

- The dynamic load capacitance (C
_{L}) estimated at 0.5 pF (C_{L}=g_{m.}T_{f}with g_{m}= ΔI/ΔV = ~ 20mA/25mV and T_{f}the transistor forward transit time = 0,8 ps) - The passive parasitic capacitance(C
_{P}) of the metal signal trace estimated at 0.5 pF - The passive parasitic inductance (L
_{P}) of the metal trace estimated at 50 pH

Under worst case conditions, the global time constant ΣT can be calculated as follows:

ΣT=Zout.C_{L} +Zout.C_{P} +L_{P}/Zout,soΣT=7.5Ω.0.5pF+7.5Ω.500fF+50pH/7.5Ω=3.75ps+3.75ps+6.66ps
= **~14 ps**

This time constant correlates well with measured DAC data of 35 ps rise and fall times (tr/tf). Furthermore, at this level, tr/tf individually represent less than 20% of the total sample clock period (of 166 ps) giving fast enough clock edges to support a first order bandwidth approximation of 10 GHz, hitting the DAC’s design goal.

Beyond this first order assessment, some special damping techniques were used to ensure dynamic stability within the DAC. Maximum overshoot (+4%) and minimal ring back (-2%) were achieved. Certainly, the fact that the B7HF200 process offers low sheet resistance, copper metallization helps further tweak and damp critical chip nodes. The performance of the resulting exceptionally clean, 6 GHz sampling is illustrated in the step response (fig 4).

**Dynamic enhancement through output pulse shaping**

Four output pulse shaping modes (NRZ, NRTZ, RTZ, RF) are provided to give system designers freedom to tailor the DAC’s dynamic response to specific output frequency bands, thus facilitating frequency planning. Most quantizer distortion can be tracked down to switching transitions. Any switching glitches are ultimately superimposed on the output signal (fig 5), however if these glitches can be removed, then the output spectral purity benefits.

This is achieved in the pulse shaped trace above, by forcing a return to zero ahead of each edge transition, visible here for both the NRTZ and RF modes. Pulse shaping is programmed via a 3-wire serial interface with two user shaping controls, the re-shaping pulse width (RPW) and pulse centering (RPB). The pulse center must coincide with the center of the transition edge if all glitch energy is to be removed. Note that this technique sacrifices a small amount of output signal power (equivalent to the area defined by RPW).

Characteristic curves (fig 6) illustrate the benefits of pulse shaping. This data shows the frequency spectrum across eight Nyquist zones out to 27 GHz (for both fs = 6 & 7 GSps) at two RPW settings (for those unfamiliar with signal aliasing read our overview here). Note that increasing sample rate noticeably expands the typical SINC (sin(x)/x) DAC output characteristic.

Up to + 12 dB improvement in harmonics levels is shown for the third harmonic due to reshaping (H3 improved from -57 dBm to -69 dBm), stretching the DAC’s ‘reach’. To further emphasize this, the following spectra (fig 7) has been produced at 6 GSps with Fout = 2940 MHz, both with (NRTZ mode) and without reshaping (NRZ mode). In NRTZ mode, the benefit of re-shaping is clearly visible.

**Measured Performance**

Output 3 dB bandwidth extends up to 7 GHz, with a guaranteed sample rate of 6 GSps enabling the generation of a 3 GHz wide, complex bandwidth. Usable output power is clearly visible in the X-band (fig 8a). The curve shows a single tone carrier at 11950 MHz with a SFDR of 50 dBc in the 4th Nyquist zone. Here the 4th harmonic dominates SFDR. This carrier frequency was carefully selected to lie on the edge of the X-band allowing many of the harmonic signals to be easily spotted as they occur in their natural harmonic order.

Increasing carrier frequency into the K-band (fig 8), with a signal reference now set at 23950 MHz in the 8th Nyquist zone, the 2nd harmonic dominates SFDR (-36.5 dBc). Again, a clean progression of harmonics is visible.

Other performance characteristics standout from these two curves, both plots show non-harmonic spurs at the mid-frequency point. These spurs can be traced back to insufficient crossover signal rejection within the DAC’s 4:1 input multiplexer. Even so, the spurs peak at -80 dBm, a very respectable level. The DAC’s noise floor is measured at close to -110 dBm.

Testing data convertor products with single or multi-tones is easily arranged in the lab. Judging the capability of a DAC on the strength of these tests, only partly tells the story. Today’s digital communication systems deploy complex modulation on large chunks of bandwidth, and so a more representative broadband test is desirable. That’s where noise o power ratio (NPR) testing helps; it exercises a DAC across a wide bandwidth and can gauge how a signal comprising of many incoherent and narrow band tones, interact and interfere with one another when mixed by the DAC. A DAC with an NPR closely matching the ideal NPR of an n-bit device is clearly a desirable broadband component.

Broadband NPR for this part is shown below (fig 9). Sampling at 7 GSps allows the generation of 3.150 GHz wide synthetic bandwidth. The resultant NPR is 42.6 dB, equivalent to an 8.6 effective number of bits (ENOB). Note the considerable NPR flatness out to 3325 MHz. The NPR test is usually performed using a digital pattern with a Gaussian noise power density. A (digital) notch filter is applied to the pattern in the frequency domain to give a ‘quiet’ zone within the bandwidth of interest. The pattern is sent to the DAC and the NPR performance is calculated as the ratio of average magnitudes of power densities measured both within and outside the notch. For an ideal DAC, the signal power within the notch is related to quantization noise only. For a real DAC, the quantization noise is combined with DAC thermal and voltage noise induced by clock jitter, (output referred) plus intermodulation products due to any cross-channel interferers.

The second NPR characteristic (fig 10) replicates a 3.150/2.700 GHz NPR pattern across + 22 GHz using 7/6 GSps sampling with the DAC operating in RF mode. This graph pair helps to highlight one of the benefits of increased sample rate. Not only does it impact the maximum bandwidth generation capability of a DAC, but it expands the SINC characteristic and the output power available in high order Nyquist zones.

**Other state-of-the-art DACs**

Texas Instruments recently described a 14-bit 8.9 GSps RF DAC using 40 nm CMOS process supporting 4G LTE applications. It features an SFDR of 50 dBc at 8.9 GSps (Fout = 4300 MHz)^{3}. Although the DAC is capable of sampling at 8.9 GSps, no measurement results are available for frequencies above 4300 MHz excluding it from most microwave bands.

Another development is Analog Devices’ 11/16-Bit, 12 GSps DACs (the AD9161/AD9162). The AD9161/62 specifies sampling rate at 12 GSps in RF mode (also called mixed mode). In RF mode, since data is inverted every half a clock period, it looks like the DAC is sampling at 12 GSps. For the EV12DS460A operating in RF (see Fig. 5), the Data inversion is not considered in the specified sampling rate (6 GSps). Therefore, the EV12DS460A and AD9161/62 are strictly equivalent in terms of sampling rate. This is validated by the fact that for all, the instantaneous bandwidth is 3 GHz.

Both Analog Devices parts feature excellent SFDR across the first two Nyquist zones of 65 dBc (Fclock = 5 GSps, Fout = 4000 MHz). Unfortunately, performance collapses at frequencies above 7500 MHz. Output power rolls off to -66 dBm (Fout = 7500 MHz) preventing them doing useful work in X and K-bands.

**Final thoughts**

With the arrival of the EV12DS460A, microwave engineers now have a practical broad-band DAC capable of projecting complex bandwidth from DC all the way into the K-band frequency. Certainly, the device is not the only giga-sample DAC available, but as shown here, it is the first one capable of projecting large synthetic bandwidth into higher Nyquist zones whilst maintaining good spectral purity. It opens a realm of exciting and innovative possibilities for new mm-wave applications.

**References**

From Infineon:^{1} J.Böck, H.Schäfer, K.Aufinger, R.Stengl, S.Boguth, R.Schreiter, M.Rest, H.Knapp, M.Wurzer, W.Perndl, T.Böttner, and T.F. Meister. “SiGe bipolar technology for automotive radar applications” in Proc. Bipolar/BiCmos Circuits and Technology Meeting (BCTM),Montreal, Canada, Sep. 2004, pp.265-268<

From Teledyne e2v:^{2} François Boré, Marc Wingender, Nicolas Chantier, Andrew Glascott-Jones, Emmanuel Dumaine, Carine Lambert, Sergio Calais. “3 GS/s 7GHz BW 12 Bit MuxDAC for Direct Microwave Signal Generation over L, S or C Bands” in Proc. COMCAS, Tel Aviv, Nov 2011

From TI:^{3} Ravinuthula, V., Bright, W., Weaver, M., Maclean, K., Kaylor, S., Balasubramanian, S., ... & Dwobeng, E. (2016, June). A 14-bit 8.9 GS/s RF DAC in 40nm CMOS achieving> 71dBc LTE ACPR at 2.9 GHz. In VLSI Circuits (VLSI-Circuits), 2016 IEEE Symposium on (pp. 1-2). IEEE.