|
100 MHz-6.5 GHz Wideband Sweep Source Module (2/2): Components, Matching, ProductionPart 2 of 2. Part 1 covers why a 65:1 span is the hard part, how the requirement was decomposed, and the dual-conversion architecture. 4. Component selection4.1 Final stageRequirement: P1dB +22 to +23 dBm, gain ≥ 12 dB, 12 V single supply, 100–6500 MHz, efficiency as high as possible. A 0.25 µm GaAs pHEMT driver in the GM1309 class meets the low band comfortably: 30–3000 MHz, +31 dBm saturated output, 17 dB gain at +30 dBm output, 50% PAE, 12 V single supply, 131 mA quiescent, 4 × 4 mm QFN, input return loss better than −15 dB, +22 dBm maximum input, −55 to +125 °C. Single supply, minimal external parts, good efficiency — the best cost-performance point in this design. It stops at 3 GHz. Two options cover 3–6.5 GHz:
Recommendation: band-split. The module already segments filtering and LO generation, so an extra switch path is not a system-level burden, and the efficiency is converted directly into thermal margin. Scaling path, for when a customer asks for 10 W, 100 W or more: 30–2500 MHz / +40 dBm / 52% / 28 V and 30–2700 MHz / +37 dBm parts in the same family, then GaN — 20–3000 MHz, 28 V, 100 W, 19.5 dB gain, 70% efficiency, or internally matched broadband GaN modules. 4.2 Gain stage and LNAA 20–6000 MHz GaAs pHEMT MMIC with 0.7 dB typical noise figure, +20 dBm output P1dB, 5 V single supply, 50 mA, internally matched to 50 Ω, available as die, SOT-89 and 2 × 2 mm LGA, serves two roles: the 15 dB transmit gain stage, and — in a dual-channel or receive-capable variant — the LNA. Internal 50 Ω matching does not mean the layout is free. DC blocks at the input, decoupling placement and ground-plane integrity still decide whether it oscillates near 6.5 GHz. On three boards using this part, two showed a 3–4 dB gain peak near 5.8 GHz that was only removed by relocating the decoupling capacitors. Selection criteria for the LNA class: in-band gain flatness (over temperature, not typical), P1dB (whether it can serve as a transmit pre-driver), and PSRR (DC/DC ripple modulates directly onto the RF). 4.3 Switches: pick two of threeLow insertion loss, high isolation, high power handling — choose two. Silicon switches give 0.5–0.8 dB loss, fast switching and low control levels, but limited power handling (P0.1dB typically +20 to +25 dBm); they belong in small-signal positions after the DSA and around the mixer. GaAs switches handle +30 dBm and isolate well, at 1–1.5 dB loss and with a negative or TTL-converted control; they belong at the output and in any T/R path. A practical note on cascading: isolation does not simply add. With more than about 10 mm of line between two switch stages, radiation and coupling along that line typically cost 5–8 dB of measured isolation versus calculation. Place cascaded switches back-to-back, keep the interconnect under 3 mm and ground it on both sides. 4.4 Filter bankFive segments: 100–500, 500–1500, 1500–3000, 3000–4500, 4500–6500 MHz. The low segments have large relative bandwidth (5:1) — a cascade of lumped high-pass and low-pass sections is cheaper and smaller than a bandpass. The high segments are narrow in relative terms (1.4:1) — coupled-resonator or ceramic dielectric bandpass. The hard rule: above 3 GHz, wire-wound inductors are past self-resonance; do not use lumped elements there. Use distributed structures — parallel-coupled lines, hairpin, interdigital — or ceramic dielectric filters. Four numbers decide a filter: in-band loss, skirt steepness, stopband rejection, and group-delay variation. For a sweep source the last one matters most: group-delay variation sets pulse broadening and sweep transient tailing. For noise-modulated jamming waveforms it is critical — above about 5 ns of variation the noise bandwidth is visibly sculpted and the effectiveness drops. Huashi supplies custom filter banks from its own thin-film and micro-assembly lines, which for small quantities and fast turnaround usually beats a large foundry's production queue. 4.5 DAC and clockA 1 GSPS, 14-bit DAC is the baseline. Companion 16-bit 500 MSPS DAC and 16-bit 250 MSPS ADC devices from the same family keep the BOM single-sourced. A cheaper route is a 12-bit 2.5 GSPS part, at the cost of 60–65 dB SFDR — the spur floor then has to be corrected in the digital domain. A more elegant option is a DAC with an integrated DDS: 1 GSPS, 14-bit, serial I/Q input, > 80 dBc narrowband SFDR. It removes the NCO from the FPGA, cutting both logic and power — attractive against a 6 W budget. Channel count: 200 MHz instantaneous bandwidth as a single real signal fits one 1 GSPS DAC. For I/Q — required for single-sideband operation and image rejection — use two DACs or one dual-channel device. The clock is the most expensive component and the one that cannot be economised. DAC clock phase noise sets output phase noise and the spurious floor. A 100 MHz TCXO (±0.5 ppm, −40 to +85 °C) as reference, multiplied to 1 GHz by a Huashi clock generator or PLL, gives integrated jitter in the 100 fs region. Three traps:
4.6 FPGA
Three FPGA pitfalls, all of them common:
4.7 Detection, monitoring and interfaceA log detector with 50 dB of ±3 dB range (40 dB of ±1 dB), +10 dBm maximum input, 16–18 mV/dB slope, 3.3–5.5 V supply, 20 ns rise/fall and a temperature-compensation pin closes the levelling loop. With −20 dB coupling the detector sits mid-range at full power and at −43 dBm at minimum output — below the ±1 dB segment but inside ±3 dB. Closed-loop levelling depends on slope and repeatability; absolute accuracy comes from the factory calibration table, which is why the low-power end still holds. RS-422, five wires (Y/Z out, B/A in, GND), through a Huashi transceiver. For cascading, the host connects to the first unit's RS-422-IN and each OUT feeds the next IN. Fit a 120 Ω termination footprint at every input; populate it only on the last unit. Without this, three or more cascaded units begin to error. Alarm and enable lines (current, temperature, PA enable) are 0/3.3 V and connect directly to FPGA I/O — with a 100 Ω series resistor and 100 pF to ground on each, otherwise internal switching noise is injected straight into the digital ground. Two 0–3.3 V ADC inputs are enough for temperature, detected voltage and supply current; a small external multi-channel ADC keeps the DAC's digital resources free. 4.8 Power12–28 V input, with 5 V and 3.3 V for the PA, DAC and clock, and 1.0 V core plus 1.8 V/3.3 V I/O for the FPGA. Two stages: a wide-input buck from 12–28 V to 5 V, then 5 V to 3.3/1.8/1.0 V. Wide-input bucks covering 4–36 V in and 2–3 A out are available from Huashi at roughly half the imported price. Three cautions:
Efficiency sets the standby figure. 90% at the first stage and 85% at the second gives about 76% overall: a 3.5 W load draws 4.6 W, i.e. 0.38 A at 12 V — inside 0.5 A. At 70% overall the same load needs 5 W and 0.42 A: still passing, with 0.08 A left. Every 5 percentage points of efficiency is a few hundred milliamps of margin. 4.9 Passives and connectorsAt 6.5 GHz, boards fail on passives more often than on ICs. Capacitors, three classes. RF coupling and matching: high-Q microwave porcelain, 0402, 0.3–100 pF, C0G/NPO, Q > 100 at 1 GHz. Huashi microwave porcelain series cost a third to a half of the established US parts and are adequate from 1 to 6 GHz. Mid-band and bias decoupling: X7R 0402/0603 — and remember the DC bias derating; a 100 nF 0603 at 5 V is effectively about 60 nF, so decouple against the real value. Bulk: tantalum or polymer, 10–100 µF. Inductors: specify self-resonant frequency, not inductance. A 1 nH wire-wound part self-resonates at 6–10 GHz, 2.2 nH at 4–6 GHz, and anything above 4.7 nH is capacitive by 3 GHz. Above 4 GHz, replace lumped inductors with high-impedance line: a 0.25 mm × 3 mm trace is about 2 nH with no self-resonance and better unit-to-unit consistency. Connectors. SMA is adequate to 6.5 GHz over this span, but use RF-grade parts — beryllium-copper gold-plated centre pin, PTFE dielectric — not the low-cost video-grade SMA, whose VSWR reaches 1.5 above 3 GHz. 2.92 mm is the option if the band is later extended. For DC and I/O, use microminiature rectangular connectors rather than pin headers; headers work loose under vibration. 5. Matching: measured comparison of three topologiesThree matching networks were built on the same device, in the same housing, with nothing else changed, and measured at seven frequencies. The absolute numbers are indicative; the trend is the point.
Topology A collapses at the top of the band because the inductors are past self-resonance by 4 GHz and the network is no longer the network that was simulated. Topology C costs 0.4 dB of extra loss and 6 mm of board length, and buys ±0.6 dB flatness. Production chose C. The 0.4 dB is recovered with 0.4 dB of front-end gain; the flatness saves the manual equalisation of every board on the line. That is the production trade: board area and a little gain, in exchange for test time. 5.1 Tuning procedure
5.2 Three pitfalls that look correctPitfall 1 — too many sections in simulation. More sections mean wider theoretical bandwidth, but each carries loss and manufacturing tolerance. Above roughly ten sections, a real board is almost certain to measure worse than the simulation. Validate with Monte Carlo on component tolerance and on etch variation before committing. Pitfall 2 — lumped elements above 4 GHz. The self-resonance table in §4.9 is not a guideline, it is a limit. A three-section lumped match that looks perfect at 2 GHz will show a 3 dB notch at 6 GHz because the series inductor has become a capacitor. Replace with high-impedance line and re-optimise; the layout is larger but repeatable. Pitfall 3 — no de-embedding. The SMA launch, the transition via and the connector pin add 0.3–0.8 dB of loss and tens of degrees of phase at 6.5 GHz. If the fixture is not de-embedded, the matching network is optimised for the fixture rather than for the device, and every board built from that design inherits the error. Measure a through line of the same length, subtract it, and only then tune. 6. Build or buyTwo parts of this module are usually not worth building in-house at low volume. The converter chain. Two mixers, two LOs, filters and amplifiers in a few tens of millimetres will absorb half the project's debugging time. Off-the-shelf up/downconverter modules with IF, RF, LO, DC and serial interfaces remove that work entirely. The dividing line: above roughly 1,000 units per year, or with unusual band and bandwidth requirements, build it; at a few hundred units with a schedule to hit, buy it and put the engineering into the PA and system integration. The power stage. Below 10 W, building from discrete or MMIC devices gives the best cost and volume. At 50 W and above — particularly for space or wide temperature — buy an internally matched module or have it assembled as a SiP by Huashi, whose in-house thin-film and eutectic die-attach lines deliver with full-temperature data. The principle: what you save is engineering hours, not component cost. A module costs more than its constituent ICs, but two months of debugging and one failed board spin buy several hundred of them. 7. SummarySix decisions determine whether a 65:1 sweep source reaches production:
About HuashiHuashi Digital Technology (Shenzhen) Co., Ltd. designs and manufactures RF, microwave and satellite communication equipment, including the HS-PA series power amplifiers (300 MHz–6 GHz) and wideband converter modules (0.8–18 GHz). Modules are built on Huashi bills of material, with full-temperature test data, and can be supplied with space-grade screening and radiation documentation on request. Contact
Specifications in this note are design references. Final performance is confirmed by measurement on the delivered unit. |