One-line mental model

Spectrum placement is set by fs / interpolation / NCO / RTZ; spur level is set by static + dynamic errors; what survives to the antenna is set by reconstruction + bandpass window.

Direct-RF DAC vs DAC + analog IQ upconversion (engineering consequences)

• Signal path complexity: Direct-RF translates and combines carriers in digital (DUC/NCO + interpolation), reducing analog blocks. Analog IQ adds LO distribution, I/Q imbalance, LO leakage, and extra calibration surfaces.
• Phase coherence at scale: Direct-RF makes deterministic multi-channel coherence easier with shared clocks and repeatable digital paths. Analog IQ often shifts coherence risk to LO phase drift and per-channel analog mismatch.
• Where images and spurs come from: Direct-RF is dominated by Nyquist-zone planning, interpolation images, clock-induced skirts, and code-dependent spurs. Analog IQ is often dominated by I/Q imbalance, LO artifacts, and mixer products.
• Clock / LO sensitivity: Direct-RF ties RF SNR and near-carrier noise tightly to the sampling clock. Analog IQ ties spectral purity heavily to LO phase noise and quadrature accuracy.
• Filtering strategy: Direct-RF uses a reconstruction/bandpass window to keep the intended zone and reject images. Analog IQ typically stacks IF/RF filtering and LO isolation across more analog stages.

How to use this page (fast reading path)

1. Start with frequency planning: choose the Nyquist zone, interpolation, and NCO placement so images land where filters can reject them.
2. Then evaluate RTZ vs NRZ: treat output mode as a spectrum-shaping knob, not as a checkbox.
3. Translate clock jitter / phase noise into RF SNR and near-carrier skirts; budget clocks early before hardware locks in.
4. Classify spurs by signature (harmonic / image / skirt / coupling) and apply the right mitigation knob first.
5. Validate multi-channel coherence and production test with short, repeatable measurements that isolate failure modes.

Architecture stack (what each layer controls)

• NCO/DUC (digital translation): places carriers and multi-tone content in the target band; sets frequency agility and multi-carrier flexibility.
• Interpolation filters: push images away from the passband; shape stopband energy so reconstruction filtering becomes feasible.
• DAC core (typically current-steering): sets the baseline INL/DNL and dynamic distortion ceiling that limits achievable SFDR.
• Output mode (NRZ/RTZ): changes the effective pulse shape and high-frequency spectral envelope; trades output power vs image behavior and wideband purity.
• Output network + bandpass window: selects the intended zone and suppresses out-of-band images/spurs; turns spectrum planning into a realizable RF interface.

When an RF DAC is the right tool (selection triggers)

• High RF band and wide bandwidth: the intended band sits in a higher Nyquist zone and needs images pushed far enough for practical filtering.
• Strict spurious mask: spectral purity must be maintained across temperature and production variation; spur mechanisms need controllable knobs (interp/RTZ/dither hooks).
• Multi-channel coherence: phased-array or MIMO transmitters require deterministic phase alignment and repeatable latency behavior across channels.
• Compute / power constraints: on-chip DUC and interpolation reduce FPGA load and simplify the digital pipeline while keeping agility.

Key terms (quick alignment)

• Nyquist DAC: typically targets baseband/low IF and relies on external upconversion.
• RF DAC: can place energy directly into higher zones with built-in DUC/NCO and interpolation.
• NCO: digital LO for frequency translation.
• DUC: digital upconversion chain (mixing + filtering).
• Interpolation: increases effective sample rate and pushes images outward.
• RTZ: output mode that returns to zero between codes, changing the spectral envelope.

Nyquist zones in one minute (why mapping matters)

• fs sets the grid: the spectrum repeats every fs. Choosing a zone is choosing which repeated copy will be kept by the analog bandpass window.
• Images are predictable: the dominant image bands typically appear around fs ± f (and around the effective sampling rate after interpolation).
• Filtering is a window: the reconstruction network does not “fix” spurs; it keeps one band and rejects the rest. Planning is about making that window realistic.

4-step frequency plan (inputs → decisions → outputs)

1. Define the target band (input).
   Specify center frequency and occupied bandwidth, plus the spectral rule that must be met (spur mask, adjacent-band limits, or a keep/reject window). Output: a clear keep band [f0 − BW/2, f0 + BW/2].
2. Choose the output zone (decision).
   Place the keep band inside a Nyquist zone where the first image bands do not overlap the keep window. Prefer placements with comfortable guard bands to the zone edges.
3. Select interpolation and NCO placement (decision).
   Use interpolation to increase image spacing and use NCO/DUC to land the content into the chosen zone. Output: interpolation factor and a feasible NCO range that preserves guard band on both sides of the keep band.
4. Predict image landing and write the filter window (output).
   Identify the first image band that must be rejected, then define passband and stopband windows (with margin) as acceptance criteria for the reconstruction/BPF network.

Common pitfalls (symptom → likely cause → first action)

• Images sit too close to the intended band → the keep band is near a zone edge or image spacing is small → change zone or increase interpolation before chasing SFDR numbers.
• Filter looks “impossible” on paper → passband/stopband margins were not defined early → write the window first, then adjust NCO/interp to create transition room.
• A spur appears exactly at fs ± f → image/alias landing as planned, but not rejected enough → verify BPF stopband window and output matching before re-tuning the digital chain.
• Multi-carrier edges fail the mask → occupied bandwidth was underestimated or guard band is too small → re-plan using worst-case BW including crest and allocation.
• Measured spurs do not match simulations → test chain reflections/overload create false tones → add attenuation/isolation and validate instrument dynamic range.
• A “good datasheet” still fails system limits → spectrum placement and windows were never locked → freeze the zone + window first, then optimize clocks and spur mechanisms.

What interpolation changes (benefit → cost → verify)

• Benefit: pushes the dominant image bands outward, creating larger transition room between the keep window and the first image.
  Verify: confirm the first image sits fully inside the reject window across the full occupied BW.
• Benefit: improves practical filter feasibility by reducing how aggressive the analog bandpass must be.
  Verify: check passband ripple and group-delay flatness versus the system limits (mask, EVM, adjacent-band leakage).
• Cost: adds digital filter delay and configuration complexity; wideband modulation becomes more sensitive to passband ripple and delay ripple.
  Verify: validate with a modulation-representative test (multi-tone or VSA EVM/ACLR) rather than single-tone only.

Quick guide: x2 / x4 / x8 (directional tradeoffs)

• Interp x2: a practical first step to gain guard band; often enough when the keep window already has reasonable spacing. Risk: modest delay; verify that the first image band clears the reject window over BW.
• Interp x4: substantially increases image spacing and relaxes analog transitions. Risk: more delay and more sensitivity to passband ripple; verify delay ripple and adjacent-band behavior.
• Interp x8: maximizes spacing when masks are tight and zones are high. Risk: higher implementation burden and more configuration states; verify across temperature and corner settings, not one lab point.

On-chip DUC checklist (what to confirm before committing)

• Configurability: available interpolation factors, NCO range/resolution, and whether response profiles are fixed or selectable.
• Repeatability: whether frequency planning stays stable across modes (same placement, same images, same band edges).
• Spur behavior: whether NCO placement or mode changes create consistent spur signatures that can be avoided by planning.
• Validation approach: confirm with both single-tone (images) and representative wideband tests (mask, ACLR/EVM) because interpolation can shift what dominates.

What RTZ changes (pulse shape → envelope → window)

• Pulse shape: NRZ holds level across the full sample; RTZ returns toward zero within the sample period.
• Spectral envelope: the pulse shape sets how energy rolls off with frequency, affecting how “friendly” the placement is in higher zones.
• Filter window: changing the envelope can move the practical boundary between what can be kept and what can be rejected by the reconstruction/BPF network.

When RTZ is often worth evaluating (practical triggers)

• Higher Nyquist zones: the intended band sits in a higher zone where images are harder to window with realistic transitions.
• Wide occupied bandwidth: multi-carrier or wide modulation needs better control of the envelope and image spacing to protect the mask.
• Tight filter windows: the reconstruction/BPF stopband has limited room; RTZ can sometimes make the keep/reject window more achievable.
• Power margin exists: the link can tolerate output amplitude changes without violating the transmit power budget.

Tradeoffs (what changes) and acceptance checks (what to measure)

• Output amplitude / power: RTZ can reduce effective output power. Check: verify the transmit gain budget still closes with the same output network and calibration.
• Noise and distortion distribution: the envelope change does not remove mismatch spurs, but it can change what dominates in a given zone. Check: compare single-tone SFDR and multi-tone/ACLR (or mask) because the “winner” can differ.
• Filter window dependency: a BPF tuned for NRZ may not be optimal for RTZ in higher zones. Check: keep the same BPF and measurement setup when comparing modes to isolate the mode effect.
• Load/matching sensitivity: reflections can change spur visibility and mask margins. Check: enforce stable test conditions (attenuation/isolation) before concluding RTZ is better or worse.

Decision card (quick rule)

• High zone + wideband + tight window: prioritize evaluating RTZ early because envelope behavior can decide feasibility.
• Lower frequency + narrowband + relaxed window: NRZ is often sufficient and can preserve output power margin.

Jitter → SNR risk (high-frequency sensitivity)

• Higher output frequency: the same jitter creates larger phase error, so SNR can collapse faster in higher zones.
• Wideband planning: jitter-driven noise can dominate total wideband noise power even when single-tone SFDR looks fine.
• Acceptance view: look for overall noise-floor lift and SNR reduction when moving the tone upward in frequency.

Phase noise → ACLR/mask risk (near-carrier skirts)

• Skirts around the carrier: energy spreads near the main tone and leaks into adjacent bands.
• Adjacent-band limits: even with good far-out noise, close-in skirts can break ACLR or spectral masks.
• Acceptance view: inspect close-in phase-noise region and integrate within the adjacent-band bandwidth.

Budget workflow (target → dominant mechanism → clock interface points)

1. Lock the band plan: center frequency, occupied bandwidth, and chosen zone/interpolation from the frequency plan.
2. Lock the RF limits: SNR/EVM constraints and adjacent-band/mask constraints that define what “good enough” means.
3. Decide the dominant risk: if high-frequency SNR is the limiter, tighten jitter; if adjacent-band leakage is the limiter, tighten phase noise.
4. Define clock interface points: reference source quality, distribution isolation, and any jitter-cleaning stage required to meet the dominant limit consistently.

Measurement notes (how to read the spectrum)

• Jitter-dominant look: the far-out noise floor lifts and SNR degrades more when the tone is moved higher in frequency.
• Phase-noise-dominant look: close-in skirts rise around the carrier and push energy into adjacent bands.
• Test hygiene: keep attenuation/isolation and reference locking stable to avoid false skirts and instrument-limited results.

Typical connection topologies (short, practical patterns)

• Diff out → Balun/Xfmr → 50Ω → BPF → Port: a common path to convert differential to single-ended while keeping RF practice.
• Diff out → Diff match → Diff BPF → Diff chain: keeps common-mode controlled and supports differential PA/measurement paths.
• Diff out → Atten/Isolator → BPF → 50Ω: increases stability and measurement repeatability when reflections make spurs inconsistent.

Stability and drive risks (what breaks when the network is not controlled)

• Poor match (S11): can create amplitude ripple, spur variability, and inconsistent results across cables and fixtures.
• Transformer/balun behavior: common-mode conversion and bandwidth/phase non-idealities can change even-order distortion and wideband mask margin.
• Reconstruction window side effects: a filter can suppress images, but passband ripple and group-delay ripple can reduce modulation quality and adjacent-band compliance.

Acceptance checklist (minimum set for repeatable RF behavior)

• S11 / return loss: stable within the keep band; verify corners (temperature, power, fixtures).
• Amplitude flatness: keep-band ripple stays within the system budget for the intended modulation and bandwidth.
• Group delay / phase flatness: verify wideband behavior, not only single-tone SFDR, because masks can fail at the band edges.
• Even-order distortion sensitivity: check 2nd-harmonic trend when converting differential to single-ended (common-mode effects).
• Repeatability: enforce a fixed attenuator/coupler strategy when spurs vary with cables or load reflections.

Knob map (what changes what)

• RTZ/NRZ, interpolation, digital dither: mostly changes code-related spur visibility and image placement.
• Clock isolation/cleaning: targets close-in skirts and clock-imprinted tones.
• Output network and partitioning: targets repeatability, reflection-driven variability, and supply/return coupling.

Fault-tree cards (symptom → evidence → first action)

Three pillars of coherence (what must be true)

• Common reference + low-skew distribution: one phase reference, a controlled clock tree, and measurable skew between channels.
• Sync event + deterministic latency: a shared trigger/SYSREF concept that makes alignment repeatable across runs and power cycles.
• Calibration hooks: a way to measure and correct amplitude/phase mismatch and drift (temperature, aging, and board-to-board offsets).

Verification card (how to prove coherence in hardware)

• Skew measurement: drive the same stimulus through all channels and measure channel-to-channel timing/phase difference at a defined test point.
• Power-cycle repeatability: repeat cold boots and confirm the alignment does not randomly “jump” to new offsets.
• Temperature drift tracking: sweep temperature and log phase offset trends to confirm drift is measurable and correctable.
• Calibration effectiveness: apply the correction loop and verify phase spread shrinks and stays bounded over time.

Board-to-board / chassis coherence (practical notes)

• Distribute reference and trigger: use a shared reference input and a shared trigger edge so every board sees the same alignment intent.
• Control path delays: treat cables and backplane routes as part of the timing budget and keep them measurable and repeatable.
• Keep calibration hooks: provide a loopback/coupler path or a pilot-based method so installation variance can be corrected in-system.

Mapping checklist (spec → system symptom → KPI risk)

• Noise floor / SNR: raises wideband noise → EVM degrades and the constellation spreads.
• Clock jitter / phase noise: adds phase-driven error and skirts → EVM + ACLR risk increases.
• SFDR / discrete spurs: narrow tones violate limits → spurious mask can fail even when noise looks fine.
• Group delay / passband ripple: wideband distortion and edge degradation → ACLR/EVM margin shrinks.
• Channel amplitude/phase mismatch: coherence breaks in arrays → beam error and sidelobes increase.

Budget steps (target KPI → mechanisms → owners → acceptance tests)

1. Freeze waveform and limits: occupied bandwidth, carriers, and mask/ACLR/EVM targets define what matters.
2. Choose the dominant KPI: EVM-driven, ACLR-driven, mask-driven, or array-driven budgets behave differently.
3. Allocate to mechanisms: noise, jitter/skirts, discrete spurs, ripple/group delay, and mismatch each get a share.
4. Assign owners: DAC core, clock tree, output network, layout/partition, and calibration each own their mechanism.
5. Define acceptance tests: pick the shortest test per mechanism (single-tone, multi-tone, modulated, temperature drift, coherence).

Fast triage (symptom → likely domain)

• Image location looks “wrong”: frequency plan (fs, NCO/DUC, interpolation, zone, reconstruction window).
• Close-in skirts dominate: clock phase noise / reference locking / clock distribution isolation.
• Fixed spur correlates with IO/supply activity: coupling via layout, return paths, partitioning, and supply filtering.
• Harmonics grow rapidly with level: output network compression/symmetry, matching, transformer/balun behavior.

Test checklist cards (purpose → setup → pass/fail → what it reveals)

Fixture and measurement chain rules (do these or results are misleading)

• Protect analyzer linearity: insert fixed attenuation before the analyzer and avoid front-end overload.
• Lock the reference: use the same 10 MHz reference for the DUT and the analyzer when evaluating skirts and EVM/ACLR.
• Make the chain repeatable: keep the same attenuator/coupler/cable set across tests; mark reference points and calibration points.
• Control reflections: use proper 50Ω terminations and isolation where needed to prevent spur variability caused by mismatch.

Practical bench BOM (example part numbers for repeatable fixtures)

Note: part numbers are practical examples; choose frequency range, power rating, and connector gender that match the target band and fixture mechanics.