|
100W-Class 8-12 GHz Broadband RF Power Amplifier Module: From Specification to Volume Production - Complete Engineering Handbook100W-Class 8-12 GHz Broadband RF Power Amplifier Module: From a One-Page Specification to Stable Volume ProductionShenzhen Huashi Digital Technology Co., Ltd. (深圳市华视数码科技有限公司) independently develops, manufactures and sells broadband RF power amplifier modules for wideband RF jamming/counter-UAS front ends, test and measurement, EMC test systems and satellite/near-Earth platform payload front ends. This handbook distills our engineering team's field experience with a 100W-class 8-12 GHz broadband RF power module into a complete walk-through - from a one-page specification all the way to stable volume production. It is written the way a chief RF engineer mentors a new hire: every number is followed by the design decision behind it. Target specification: 8-12 GHz, saturated output >100 W, gain 50 dB, in-band flatness ±1 dB, harmonics -50 dBc, spurious -60 dBc, efficiency 22%-30%, 28 V supply, input return loss ≤-10 dB, 5:1 load VSWR without damage, over-temperature shutdown at 85 °C, over-current shutdown at 20 A, 160×120×25 mm, 2 kg, SMA input / N output, hybrid D-type power connector.
1. Scope FirstBefore anything else, define the boundary: this link is built only for broadband RF countermeasures, test and measurement, EMC testing and satellite/near-Earth platform payload front ends. It performs no directed action against any specific communication standard. In essence it is a broadband, clean, reliable RF power source. Device selection, band planning and filtering strategy all live inside this box. 2. Simulation: Only Three Things, Each With Concrete Settings2.1 Three layers of simulation, on a fixed scheduleLink-budget simulation - a spreadsheet, done in half a day: gain, power, compression points, noise figure and spurious budget in one pass. With 0 dBm input and 50 dB total gain, the final-stage input already sits at +50 dBm; any 0.5 dB of extra insertion loss earlier in the chain demands 0.5 dB more drive from the final stage. Change one number in the sheet and you see the global impact. Circuit simulation - harmonic balance plus load-pull with the vendor's large-signal model, one week. 3D EM simulation - only four items: the cavity, connector transitions, the filter, and the final-stage high-current traces. Nothing else. Newcomers fall into two extremes: simulating nothing and relying on tuning to save the day, or putting every screw into a 3D model and still not having a board three months later. 2.2 Harmonic-balance settings (generic commercial software)Harmonic order 5, MaxOrder 7 - enough to resolve the third harmonic with margin. Adaptive sampling, 8-12 GHz in 200 MHz steps (61 points). A four-way-combined large-signal simulation takes about 20-40 minutes on an ordinary workstation. Before running, do one thing: overlay the model's small-signal S-parameters on the datasheet curves. If S21 differs by more than 0.5 dB or S11 phase by more than 10 degrees anywhere in 8-12 GHz, do not run large-signal yet - check the model version, whether package parasitics are included, and whether the bias conditions match the datasheet. Five minutes here saves two days of "the result is wrong but I don't know which part is wrong." Load-pull: run one sweep each at 9, 10 and 11 GHz. On the Smith chart, the power and efficiency contours should not coincide in the same region. The power-optimum point usually sits on the capacitive side of the efficiency point; take the intersection of the 0.5 dB power contour with the efficiency contours as the matching target. If the target impedances at the three frequencies form a smooth arc on the Smith chart, the band is matchable; if the points jump around, split the band into two matched segments or change the device. Experience discount: derate simulated power by 15% and efficiency by 10% for expectations. If simulation says 140 W and the prototype delivers 120 W, the model is good; if simulation says 33% efficiency and you measure 30%, that is normal. Never report simulated values as guaranteed specifications. 2.3 Thermal simulation: junction temperature, jointly with EMBuild the heat-flow model from the junction up: die → eutectic layer (AuSn, assume 3% void rate) → CuW carrier → thermal interface → cavity base → cold plate, filling every layer with measured or handbook values - never software defaults. Interface materials are the most under-estimated item: a 0.2 mm thermal pad over 1 cm² contributes more than 0.5 K/W; a 0.05 mm indium foil under proper pressure drops below 0.1 K/W. Same "put a layer in between", 15 °C difference in junction temperature. Bond-wire current density, computed together: final-stage device at 40 W output, 28 V drain, 30% efficiency → die current ≈4.8 A. Four parallel 25 µm gold wires carry 1.2 A each - close to the long-term safe limit of a 25 µm gold wire (about 1.5 A). Switch to 20 µm wires at the same count and each carries 1.9 A: fine short-term, but random wire burnouts appear in production. Either add wires or switch to 50 µm aluminum ribbon. Cover resonance, computed in passing: a 100×80 mm, 2 mm-thick aluminum cover has its first bending mode near the low end of the RF band; coupled with the internal air cavity it can swallow energy at some frequency, showing up as an isolated 0.5-1 dB dip. Fix: stiffen or thicken the cover, or drill a row of 3 mm holes at 15 mm pitch to break the membrane effect - hole diameter must stay below 1/20 of the lowest operating wavelength to avoid leakage. 3. Prototype: Fabrication, Machining and AssemblyThis part is process, not technology - but it decides when your prototype reaches the bench. PCB fabrication - brief the board house on four things: laminate brand and thickness; copper weight (RF layer 1 oz minimum, power traces up to 2 oz); controlled-impedance line width and tolerance (±10%, written into the fabrication notes); and plated-hole treatment (RF ground vias must be filled and capped). Four or six layers, normally 7-10 days, 5 days expedited - do not trade impedance tolerance for two days of expedite fees. Mechanical parts: cut aluminum soft tooling first (3-5 days), dry-fit once to verify assembly relationships and the interfaces to antenna and cold plate, then cut production parts. Cavity machining accuracy is decisive for the filter: cavity dimensional tolerance ±0.05 mm, surface roughness Ra 1.6 µm or better, silver plating 5-10 µm. Machine the filter cavity and the PA cavity separately and assemble them - one out-of-tolerance dimension should not scrap an entire housing. SMT and soldering: lead-free reflow for the RF board, profile set by the solder paste. GaN devices are temperature-sensitive: peak temperature below 245 °C, time above liquidus within 30 s. Placement accuracy within 0.1 mm - an off-center die means an off-center match. For flange-packaged final-stage devices, screw-mount with thermal grease or indium foil, using a torque screwdriver (never by feel): inconsistent torque means inconsistent thermal resistance, a hidden source of production parameter spread. Bond wires 25-50 µm gold or aluminum, ultrasonic bonding, with pull-force sampling inspection. Assembly order: cavity cleaning (ultrasonic, deionized water, dry) → cold plate and temperature sensors → PA carrier → boards → filter → connectors → cover → hermeticity check (if required) → power-up. Photograph every step; in production this is your traceability record. 4. Core Component Selection: Domestic-First, With Watch Items and Substitution CautionsModel names below are verifiable public examples only; in a real project always defer to the latest datasheets, sample measurements and supply channels - and keep a second source for every critical position. Final-stage and driver GaN: CETC 13th/55th Institute ecosystem packaged devices and dies (X/Ku band, tens of watts to 100 W class, full reliability data, long lead times); Guobo Electronics' GaN PA line (LF through Ku band, strong production economics); Huatai Electronics' GaN PA family (0.1-6 GHz, 15 W drivers to 200 W final stages, ready-made two- and three-stage combinations for fast chain assembly); plus Huiruixin, Sichuan Yifeng, Junweixin, Chengchang, Xinbaite, Shidai Suxin, Suzhou Nengxun, Haiwei Huaxin, Sanan Optoelectronics - each strong in particular bands, power levels and package styles. For supply, ask every candidate for: datasheet, full-temperature data, lot-consistency data, load-pull curves and second-source status. Driver and front-end stages: plastic-package GaN small-signal power devices (10-15 W, 28 V, 0.1-6 GHz, 14-15 dB gain, >60% efficiency) are the universal driver-stage part at tens of RMB each; GaAs pHEMT gain blocks serve the small-signal section with >30 dB gain and high-VSWR survivability. LNA: GaAs broadband LNAs, NF 1-2.5 dB, gain 15-25 dB, P1dB above +10 dBm. Guobo, OCR (Zhongke Hantianxia) and several MMIC specialists have catalog parts. Filters, circulators, couplers: these are "custom-first" categories. Cavity filters and circulators are essentially all custom-made - many domestic ferrite and cavity-device vendors (CETC institutes and private companies) have mature capability, lead time 2-4 weeks: order early, and write power handling, insertion loss, directivity and temperature range into the technical agreement. Power supply: 28 V drain bus from domestic PMICs - Silergy, SGMicro, Southchip, Jedonpower all have suitable bucks and load switches; gate negative bias from an SGMicro-class negative charge pump plus LDO; sequencing via a dedicated sequencer or FPGA reset logic. Decoupling caps from Fenghua Advanced, Sunlord-class MLCC vendors (0402, C0G, 50 V); bulk energy storage in tantalum or polymer capacitors; chokes from Sunlord/Magtop wire-wound or film inductors - check exactly two numbers: SRF above the operating frequency, saturation current above twice the stage maximum current. FPGA and memory: Unisoc (Ziguang Tongchuang), Anlogic, Gowin (industrial grade, low cost, domestic tool chains); for high-reliability and space, Fudan Micro's rad-tolerant series. Configuration storage: domestic serial flash, industrial grade with write protection; capacity is not the point. Data converters and signal chain: a 12-bit, sub-Msps ADC (reading detectors, temperature, current) is sufficient - 3Peak, Novosense, Runshi, Chuantu all have suitable parts; op-amps and comparators from the same vendors. Connectors: RF connectors from CNLiTech (AVIC Optronics) or domestic microwave connector makers. For SMA and N at the 100 W level, select high-power versions specifically rated for 100 W CW - an ordinary SMA is rated only tens of watts at 10 GHz; choose extended inner conductor / enlarged air-gap types, or step up to a 2.92 mm connector. DC: hybrid D-type, domestic mil-style connector makers all supply it; check pin current rating - at 28 V / 20 A, design each power pin for at least 10 A and parallel two pins. 5. Prototype Debugging: From First Power-Up to Full ComplianceDebugging is presented strictly in order - the wrong order burns hardware. Before power: DMM the 28 V-to-ground resistance (no short), gate-bias-to-ground resistance, and every rail; confirm soft-start time; set the PA gate bias to its most negative (fully pinched-off). First power-up: control rails only (3.3 V, 1.8 V, 1.0 V) - verify FPGA configuration, current level, serial port. Then gate bias, measuring quiescent current stage by stage. Then 28 V: the PA should draw only quiescent current, total roughly 8-10 A, matching the datasheet quiescent figure. Small-signal sweep: full two-port VNA calibration (cal kit rated past 12 GHz), measure S21 and return loss, examine gain-curve shape. Goal of this round is "right shape", not absolute values. If the shape is wrong, check three things first: oscillation (broadband sweep), a bad connector transition (swap connectors), or loose cover screws. Power sweep: signal source plus power meter, from -10 dBm upward; log output power and DC current at every step; plot Pin-Pout and efficiency curves. The curve shape tells you everything: sudden flattening = saturation, normal; sudden droop = supply current limit or protection trip; a "step" = an earlier stage compressing first. Harmonics and spurious: spectrum analyzer swept to 40 GHz (third harmonic), RBW 100 kHz, VBW 300 kHz, attenuation set so the analyzer front end never overloads - the classic rookie mistake is feeding 100 W into a spectrum analyzer; insert 30-40 dB of fixed attenuation plus a variable stage. Correct harmonic readings for the attenuator's frequency response, or you will read phantom "improvement" above 12 GHz. Two-tone linearity: two sources plus a combiner, 1 MHz or 10 MHz spacing, measured at 6-9 dB back-off from rated power. The datasheet has no IM3 spec, but when customers use this as a jamming source, in-band intermodulation creates adjacent-channel "glitches" that invite complaints - so set an internal spec: IM3 better than -25 dBc at 6 dB back-off. This is tuned via bias point: more quiescent current, better linearity, worse efficiency - an unavoidable trade. My rule: set quiescent current at 80% of the linearity-optimum point, keeping efficiency around 26-27%. Pulse and broadband signals: oscilloscope on the pulse envelope (rise time, droop, overshoot); droop above 10% means insufficient supply energy storage or too-small interstage coupling capacitance. For broadband noise-like signals watch the power-meter reading and in-band flatness: high crest factor means backing off average power by 3-6 dB, or peaks drive the final stage into deep compression and average power actually falls. Temperature: thermal camera over package, carrier, cavity and cold-plate inlet/outlet; record steady-state temperatures. Then chamber tests: -20 °C and +50 °C, two hours soak each, measure gain and power change. Compute the gain temperature coefficient - if it exceeds 0.02 dB/°C, widen the ALC loop bandwidth. Protection functions: over-temperature - hot-air gun near the thermistor: shutdown at 85 °C, recovery at 65 °C; over-current - increase load (or simulate): shutdown at 20 A within 5 s, verify latching; VSWR - terminate into a mismatched load (sliding short or tuner), step from 1.5:1 to 5:1, verify power folding and shutdown, and inspect the device afterwards. All three tests are mandatory - and run them in front of the customer, because this is exactly what they fear most. 6. Volume Production: From Building One to Building One HundredThe core of production is not "every unit can be tuned into spec" but "every unit is in spec without tuning". The distance between the two is the entire content of process maturity. 6.1 Where parameter spread comes from, and how to close itThe power distribution across one hundred units usually has three sources: device lot spread (more than half), placement and bonding process spread, laminate permittivity and cavity machining spread. Close them in this order:
6.2 Calibration flowEvery board runs automatic calibration before shipment: onto the production fixture (fixed attenuator + directional coupler + power meter + spectrum analyzer), automatically measuring full-band gain, computing the attenuator compensation table written to EEPROM, measuring forward-detector power calibration curves (3 or 5 points; full-temperature only spot-checked hot/cold), measuring reverse-detector threshold, and exercising protection. Three to five minutes per board, one serial number per board, data into the database - service can trace every device. 6.3 Burn-in and screeningHigh-temperature burn-in: 50 °C ambient, full load, 48-96 h continuous, full retest before and after; any power drift beyond 0.5 dB is pulled for analysis. Temperature cycling: -20 to +50 °C, ten cycles. Vibration and shock per the customer's environment (vehicular, airborne and naval profiles differ). These three screens eliminate essentially all early-life failures. 6.4 The five most common production-line failures
6.5 Making "cost-effective" concreteCost concentrates in three buckets: final-stage GaN devices (30-40%), cavity and filter machining (20-30%), labor for tuning and test (15-20%). Three levers:
There is also a make-or-buy calculation for module-level delivery: buying a "final stage + driver + micro-assembly" module from a capable supplier looks more expensive per unit than buying bare devices, but once you charge back the micro-assembly equipment, staffing, yield-ramp time and rework risk, it usually wins in small volume. Self-building a micro-assembly line pays off only beyond several hundred units. Compute your own crossover point - but make the decision at that moment, not after volume arrives and the process is not in your hands. 7. Matching-Network Schematics: Topologies You Can Copy DirectlyEverything above is theory; this section is drawing practice. Walking from the device outward, here is each segment's topology in words - place the elements exactly like this in your schematic. Final-stage output match (four-section low-pass). From the drain pad out: drain pad → a wide, short high-impedance microstrip (0.6 mm wide, 3 mm long - a series inductance nudging the impedance inductive) → first shunt-C pad (100 pF 0402, two in parallel) → second microstrip (0.8 mm wide, 4 mm long, ~20 ohm) → second shunt-C pad (10 pF - this one handles the band top) → third microstrip (50 ohm, 6 mm, pure phase compensation) → DC blocks (two 100 pF in parallel) → a pi tunable position (three pads: series pad stuffed with 0 ohm, shunt pads empty) → output test coupler → filter. Of the four sections: section 1 sets the low end, section 2 the high end, section 3 shapes the middle, and the pi position is the only place production ever touches. Final-stage input match (band-pass, gap-filling). Gate pad → gate series resistor (10-22 ohm, tight against the pad, dedicated to suppressing low-frequency oscillation) → a short 50-ohm line → shunt C (5.6 pF) → series inductance (realized as high-impedance microstrip, ~2.7 nH) → shunt series-resonant branch (L+C to ground, resonant at 9 GHz, specifically filling the 8-10 GHz dip) → DC block → pi tunable position → from the driver stage output. Interstage match (driver to final). Driver output → DC block → a tapered transition (three steps from driver impedance to 50 ohm) → shunt resistive branch (100 ohm in series with 10 pF to ground - absorbing out-of-band only, nearly invisible in-band) → pi tunable position → final-stage input match. I put this shunt RC in every broadband interstage: in-band penalty about 0.05 dB, but out-of-band (especially above 20 GHz) rejection improves by 6-10 dB - the specific cure for "out-of-band oscillation". Bias feed (identical every stage). Drain → choke (wire-wound, 2.7-5.6 nH, saturation current > 2× stage maximum, with a NiZn bead over it) → node with three shunt capacitors (100 µF tantalum, 1 µF, 100 pF) → 28 V bus. Gate → 22 ohm resistor → node with 10 nF + 100 pF → negative LDO output → potentiometer or DAC. Harmonic handling (the second killer move). In many broadband PAs I deliberately make the output match's first shunt capacitor a "series-resonant-to-ground" form - an L+C in series to ground resonated exactly at the second harmonic (e.g. 20 GHz when operating at 10 GHz). At the fundamental it is effectively absent (far off resonance, high impedance); at the second harmonic it is a deep notch. This "second-harmonic short" buys 8-12 dB of harmonic improvement at zero insertion-loss budget. The cost is frequency sensitivity: within one octave you can only cover one segment. My split: notch over 8-10 GHz, and let the cavity filter handle 10-12 GHz. Two segments with divided labor - the standard solution for octave-band amplifiers. 8. The Frequency-Conversion Chain, Treated ProperlyStandard upconversion architecture: DDS (or FPGA + DAC) generates the IF → low-pass filter (removing DAC image and clock leakage) → VGA (the level handle of the conversion chain) → mixer RF port. LO side: PLL + VCO → band-pass filter (LO harmonics) → amplify to the mixer's required LO level (typically +7 to +13 dBm) → mixer LO port. Mixer output → band-pass filter (select upper or lower sideband, reject LO and IF leakage) → driver chain. Three numbers for mixer selection: conversion loss (passive 6-9 dB; active mixers can show gain but worse linearity); LO-RF isolation (20-35 dB - this number directly determines how much LO leakage your filter must remove); IIP3/OIP3 (keep 10 dB or more of margin, or the mixer itself becomes the spurious source). Domestic passive and active mixers from CETC institutes and MMIC specialists are mature in the 8-12 GHz double-balanced category. A concrete frequency plan: the system outputs 8-12 GHz. A single conversion with IF 1.6-2.4 GHz and a 6.4 GHz LO (upper sideband) only covers one IF-bandwidth-wide slice - covering the full 4 GHz in one shot would require a 4 GHz IF bandwidth, unrealistic. So split into two segments: low segment 8-10 GHz with IF 1-3 GHz and 7 GHz LO (upper sideband); high segment 10-12 GHz with IF 1-3 GHz and 9 GHz LO (upper sideband). Both LOs come from one PLL switched via an SPDT, or two VCOs. This "segmented conversion" idea matters: an octave of instantaneous bandwidth that one conversion cannot cover becomes two half-octaves, each easy - the price is one LO chain and one switch. Estimating LO phase noise: LO phase noise transfers to the output through mixing; output phase noise equals LO phase noise plus 20·log(f_LO/f_IF). Roughly: "the LO's phase noise lands directly on the RF". So back-derive the LO spec from the output requirement: if the customer requires output phase noise better than -90 dBc/Hz at 10 kHz offset, the LO must reach about -95. Put the PLL loop bandwidth at the crossing point of the noise curves: too narrow leaves VCO close-in noise unpressed; too wide lets in the PFD noise floor. Use the PLL vendor's own simulation tool for the loop filter - never hand-calculate, and never copy someone else's loop parameters: different PFD frequencies mean completely different parameters. Three roads spurious takes: (1) LO leakage - fought with mixer isolation plus filtering; (2) IF harmonics combining with the LO - fought by pressing IF harmonics with the IF low-pass (IF harmonics landing in-band after mixing are the hardest to find); (3) digital clock coupling - fought with layout: separate digital and RF zones, clock traces never crossing the RF zone, separate power. Road 3 is the most overlooked: an FPGA clock coupling through a supply line into the LO produces a comb of spurs spaced at the clock frequency. When you see a comb spectrum, your first reflex is to trace the digital clock's coupling path. 9. Amplifier Operating Classes: How to ChooseNewcomers must nail this, because it governs every efficiency-vs-linearity trade. Class A: 360° conduction angle, best linearity, theoretical efficiency ceiling 50%, practically 20-30%. Used in small-signal front ends where power is small, efficiency irrelevant and cleanliness paramount. Class AB: conduction between 180° and 360°, theoretical ceiling 50-78%, practically 40-60%, acceptable linearity. The standard choice for driver and final stages. Quiescent current is the only knob: raise it toward Class A - better linearity, lower efficiency, more heat; lower it toward Class B - higher efficiency, visible crossover distortion, worse IM3. Class F and continuous-F: harmonic terminations (2nd shorted, 3rd open) square the voltage waveform and half-rectify the current; theoretical ceiling near 90%, practically 60-70% - but narrowband, because it depends on precise harmonic phasing. Pure Class F cannot span an octave; you can only borrow its idea - the "second-harmonic short" above comes from Class F, which is exactly why broadband GaN PAs run 5-8 points more efficient at the band's low end: the low end happens to allow second-harmonic treatment. Setting it in a real project: final and driver stages at Class AB, quiescent current starting at 30-60 mA per mm of total gate width; driver and front-end stages at Class A or deep AB with generous current. The system efficiency budget then reads: final stage 55-60% (device drain efficiency), driver stage 40%, front-end stages 20-30% each, plus interstage network losses → system PAE lands at 22-30%, which reconciles with the specification sheet. So when you measure only 18% system efficiency, don't panic: measure each stage's efficiency separately - nine times out of ten it is an under-biased driver stage or excessive interstage loss. 10. Circuit Details You Must Get RightGate series resistor placement: tight against the gate pad, no farther than 2 mm. Any farther and the trace between pad and resistor becomes an antenna where low-frequency oscillation starts. Value 10-47 ohm: too large sacrifices gain (it divides against the device input capacitance), too small fails to damp oscillation. Start at 22 ohm. Source grounding inductance: any inductance in the source-to-ground path introduces negative feedback, reduces gain, and on wide bandwidths makes gain rise with frequency (inductive reactance grows with frequency). So via-fence the source pads "many and close" - at least four vias, each within 0.5 mm of the pad, pitch within 1 mm. Done right, you gain a free decibel and the band top stops rising. Thermistor placement: on the final-stage device carrier - not on the cavity, not on the board. On the cavity you measure case temperature, lagging junction temperature by tens of seconds; by the time it reads 85 °C the junction has blown past 175 °C. My practice: two sensors - one on the carrier for fast protection, one on the cavity for environmental monitoring and derating control. Current-sense grounding: the sense resistor sits in the 28 V return; the differential amplifier's ground must tie to power ground, not digital ground, or switching noise contaminates the reading - symptom: "current readings jumping around" and nuisance protection trips. Add RC filtering at the diff-amp input, corner around 1 kHz, killing switching noise while keeping the 5-second delay intact - compute the time constant, keep it under 100 ms. Enable-signal routing: over-temp and over-current protection signals take dedicated traces to the FPGA and, through a hardware AND gate, pull the PA enable directly - passing through no programmable logic. The reason is plain: software hangs, hardware does not. Route this trace like a life line: short, straight, ground-guarded, far from clocks. Star grounding of supplies: every stage's 28 V taps individually from one point (the bus capacitor's positive terminal) - never daisy-chained (final stage first, then driver). Daisy-chaining lets the final stage's large current ripple modulate the driver's supply voltage, producing intermodulation - symptom: "IM3 10 dB worse than expected, varying with final-stage power". I have hit this pit at least three times, each requiring a full re-route. 11. First Power-Up: The Five Steps, and What "Pass" Looks LikeStep 1 - resistance only, no power. 28 V-to-ground should read kilo-ohms (bleeder resistors and PMIC input capacitance); under 100 ohm, hunt the short. Gate-to-ground: open or megaohms. Each low-voltage rail: hundreds of ohms to kilo-ohms. This looks like rookie work; I still do it every time, because one short costs a whole board and half a day. Step 2 - control rails, is the logic alive? FPGA configured (done LED or status register), serial port alive, current tens to hundreds of milliamps. Above 1 A means FPGA power sequencing is wrong - I/O voltage up before core - cut power immediately and fix sequencing; do not push through. Step 3 - gate bias, quiescent current per stage. This is the device physical exam. Compare each stage's quiescent current against design; adjust gate bias for deviations beyond 20% and record. Watch stability: slowly drifting current means an unstable gate regulator or self-oscillation. Also sweep a spectrum analyzer 0-20 GHz: any unexpected spur is oscillation - cure it before applying RF power, or the device burns. Step 4 - 28 V, no RF. Total current should equal the sum of stage quiescent currents plus supply/control overhead, about 8-12 A. Noticeably higher: check for a saturated choke or leaky DC block. Touch the packages - at quiescence they should be merely warm; hot means excessive quiescent current or wrong gate bias. Step 5 - small signal, sweep. Feed -20 dBm and record the gain curve: this becomes the "base map" for all later debugging - archive it and compare after every change. Canonical shapes and their causes: uniform shift - gain allocation or attenuator setting; low end sagging, top normal - input match low, or DC blocks too small; top sagging, low end normal - connector transition or oversized shunt capacitors, or laminate permittivity off from simulation; a deep mid-band notch - a cavity resonance or a shunt branch resonating in-band; overall ripple beyond ±2 dB - the match itself is wrong, go back and redo the network. 12. The Two Things Production Always Tries to Skip - and Must NotGolden sample. A production line must have one golden sample - fully tuned, fully compliant, having passed the complete environmental qualification - locked in a cabinet. At every lot change, material change or process change, build a fresh board and compare its curves against the golden sample under identical conditions. Without one you can only compare against the specification, and a specification is a line while a golden sample is a curve - curves tell you where things changed; lines only tell you pass/fail. First-article inspection. The first unit of every lot gets the full test suite - not just power: harmonics, spurious, two-tone, full temperature, protection actions. Only after the first article passes does the rest of the lot follow the simplified flow. I have watched too many "skip the first article, scrap fifty boards" episodes: save one hour, lose a week. 13. Eight Application Scenarios - Each Weighs the Specifications DifferentlyThe same hardware, but in every scenario a different specification is the lifeline. Knowing this tells you what to negotiate with the customer and what to fight with the test bench over.
After eight scenarios, one pattern: the hardware is the same, but which specification is the lifeline differs completely. So when taking a project: first, nail down the scenario; second, translate the scenario into specification priorities; third, only then draw the block diagram. Reverse the order and you discover at delivery that the one line the customer cared about most is missing. 14. The Scenarios One Level Deeper: Real Field Conditions and Specification PrioritiesScenario 1: fixed low-altitude security transmit front end - power plus survivabilityTypical requirement: a few kilometers of coverage, 24/7 unattended operation with remote monitoring. One-line spec priority: "power per the link budget, reliability per the thermal design, compliance per the spurious spec." The link budget, in a formula you can use directly: FSPL(dB) = 32.45 + 20·log(f_MHz) + 20·log(d_km). At 10 GHz and 3 km: 32.45 + 80 + 9.54 = 122 dB. Target receiver sensitivity -95 dBm; TX and RX antenna gains 10-15 dB each; required transmit power ≈ -95 + 122 - 25 = +2 dBm. Free space alone needs one watt. So why 100 W? Because the real world is not free space: buildings, trees, rain and snow, ground reflection, multipath, target aspect changes - margin alone is routinely 20-40 dB. Engineering practice: add a 30 dB "environment allowance" on top of free space, and the requirement jumps from 1 W to 100 W. The customer's 100 W is not arbitrary - it is real money buying margin. Explain this arithmetic to the customer and "can you do 200 W?" stops coming - tripling power buys only 5 dB, while a higher-gain antenna buys 6-10 dB for free: money on the antenna beats money on the die. The most frequent field problems, by frequency of occurrence:
Scenario 2: vehicular platform - vibration and supply are the two mountainsTypical conditions: ignition transients can dip the 24 V bus below 9 V; load steps spike above 60 V; vibration PSD peaks across 10-500 Hz at several g RMS; temperature -30 to +60 °C. What to design in: two-stage supply entrance protection - stage one, TVS diode plus self-recovering fuse, clipping 60 V+ spikes below 40 V; stage two, wide-input buck, 9-36 V in, 28 V out. The PA itself tolerates 40 V transiently, with devices rated 65 V or more drain - real margin. Skip this stage and the first speed bump or ignition produces "random, untraceable device burnouts" that can never be reproduced in the field - only design can guarantee against them. Structurally: machined-from-solid cavity; cover screws at 50 mm pitch; filter tuning screws staked with thread-locker plus lock washers (tune first, witness-mark, then stake with single-component silicone - never epoxy, whose cure shrinkage pulls the screw and shifts the response by fractions of a dB); no fewer than four board-fixing points in the cavity; heavy components (filters, circulators, large electrolytics) independently supported or potted. Potting material: low-modulus silicone - hard epoxy cracks solder joints under vibration and thermal cycling. A field detail: the antenna sits on the vehicle roof, the amplifier inside the cabin, joined by 10-20 m of feeder. Ten meters of low-loss feeder can eat 3 dB at 10 GHz - your 100 W arrives as 50 W. Therefore "mount the amplifier near the antenna" must be a stated premise of the vehicle proposal, or the amplifier becomes part of the antenna pedestal. Say it at proposal stage; after installation, discovering a power shortfall leaves nobody an exit. Scenario 3: man-packed platform - consumption decides mission lengthA man carries at most 15 kg; a battery above 5 kg is uncarryable. The entire design objective mutates from "power" to "effective RF joules per battery joule". The arithmetic: 100 W CW at 25% system efficiency draws 400 W DC; a 1 kg lithium pack (~250 Wh) lasts under 40 minutes. Pulsed operation at 10% duty drops average draw to 40 W - the same pack runs over six hours. The correct answer for man-packed use is pulsed/burst operation, agreed at proposal stage: the customer wants "the decisive moment delivered", not "always on". Knock-on design changes: the supply scales from 400 W continuous to 400 W pulsed / 40 W average; storage capacitance is sized by pulse width (1 ms pulse, 400 W, 2 V allowed droop: C = I·dt/dV = 14 A × 1 ms / 2 V = 7000 µF - entirely manageable); cooling drops from liquid to natural convection plus phase-change material, or simply the housing's thermal capacity over one mission; pulse rise time and droop become the critical specifications, because droop beyond 10% renders the pulse tail useless. Scenario 4: EMC-sensitive zones - spectral cleanliness is the lifelineAt airports, prisons and major-event venues, the fear is never "how far can you reach" but "where did your energy land". Civil aviation navigation and DME bands, GNSS frequencies, private-network channels - all red lines. Design requirements: out-of-band spurious below -70 dBm (10 dB stricter than the spec sheet's -60), harmonics below -60 dBc, plus spectrum self-check: a switch behind the output coupler routes to an internal wideband detector or compact spectrum-monitor channel; scan before transmit; open the final stage only when clean. Software-wise simple - but it converts "an amplifier" into "a manageable transmitter", which customers pay for. One more practical move: give every shipped unit a "factory spectrum fingerprint" - archived full spectra at several representative frequencies. When a field complaint arrives, the fingerprint proves the unit left the factory clean and settles the responsibility boundary. This has saved me more than once. Scenario 5: EMC test-system power source - accuracy is the productTest-system customers never ask how far you reach; they ask three things: is the setpoint accurate, is the power stable, does power jump when frequency changes. 80% of the product's value lives in the ALC loop and the calibration data. Targets: setpoint accuracy ±0.5 dB (full band, full temperature); short-term stability (1 min) better than 0.1 dB; long-term (8 h) better than 0.3 dB; post-frequency-switch settling under 100 ms. To get there: logarithmic detectors calibrated at every 10 MHz point with data stored in EEPROM; a dual ALC loop - fast loop (millisecond, handling pulses and fast fluctuations) trimming the VGA, slow loop (seconds, handling drift) trimming the digital attenuator; low-drift reference. Delivered, this machine becomes a "standard component" in test systems and prices above competitors. Scenario 6: satellite payload front end - the highest barThis scenario reshapes design methodology fundamentally. Device level: everything space-grade or industrial-grade with radiation data plus screening. Radiation effects split three ways - total ionizing dose (threshold shift, leakage growth; GaN and GaAs are naturally TID-robust, but silicon-based control ICs and FPGAs are the weak point), single-event effects (SEU, SEL, SEB - the PA device fears SEB most, operating at high field and high current), displacement damage (hard on optoelectronics, mild on RF devices). Against single-event burnout: derain drain voltage below 60% of rating, series gate resistor, small series drain inductance or an e-fuse for fast interruption. This is a hard requirement - SEB is permanent, with no recovery. Material level: outgassing is a hard line. TML < 1%, CVCM < 0.1% - by this standard ordinary potting compounds, heat-shrink tubing, wire insulation, even some label stocks are disqualified. Use space-grade silicones, polyimide wire, PTFE insulation. Solder: SnPb or whisker evaluation. Process level: full traceability - device lot records, per-step operator/equipment/parameter records, approved rework. Domestic space-grade micro-assembly lines run exactly this system, which is why supplier selection should weigh the quality system, not just the devices. Test level: thermal vacuum (full spec sweep at several temperature plateaus - in vacuum there is no convection, so the conduction-radiation thermal path is designed completely differently from atmospheric), thermal cycling (-55 to +85 °C, hundreds to thousands of cycles, watching solder and bond-wire fatigue), random and sine vibration, mechanical shock, radiation (TID and SEE), EMC. The whole suite consumes 6-12 months - schedule space projects accordingly. Scenario 7: phased-array final-stage unit - consistency is everythingArray channel counts run tens to hundreds; per-channel amplitude and phase consistency directly set beam-pointing accuracy and sidelobe level. Rule of thumb: amplitude consistency better than ±0.5 dB and phase better than ±5° to hold sidelobes below -25 dB. The design philosophy here differs from all previous scenarios: not "tunable", but "needs no tuning". Every module's components, trace lengths, device lots and assembly process identical - consistency by process repeatability, not hand adjustment. The trade: spend more on material (tighter-tolerance capacitors, finer cavity machining) to convert "adjustable" into "non-adjustable" - because adjustable means every unit gets tuned, and across hundreds of units consistency always spreads. Concrete measures: fixed-element matching networks with no tunable positions (or populated with fixed values in production); trace lengths fixed by design rule, hand-wired jumpers forbidden; device lots locked for the whole build; automated placement and bonding with digitized process parameters; automatic test writing calibration data per module, residual phase/amplitude error compensated in the digital domain. Walk this road and your array is produced "like an industrial product" rather than crafted "like an artwork" - the difference between winning volume orders and not. Scenario 8: instrument-grade PA units - repeatability is the brandThese customers buy "every unit's curve is the same". The highest-value investments are calibration methodology, temperature-compensation algorithms and aging models. Calibration decides whether factory data is credible; compensation decides cross-environment consistency; aging models decide whether the customer returns for recalibration in year three. A concrete practice: on the production line, five-point power calibration per frequency across three temperatures, fitted into a two-dimensional surface (frequency × temperature) stored in the unit; at use time the machine interpolates compensation from live temperature. Done, two units overlaid agree within 0.2 dB - exactly what an instrument should look like. 15. Three Cross-Scenario Field Rules, With NumbersRule 1: keep an "unforeseen loss" line in the link budget - 3 dB, non-negotiable. Reference split: cavity filter 1.0-1.5 dB, RF switch 0.5-0.8 dB, two semi-rigid cable runs 0.3 dB, four connector pairs 0.2 dB, combining/splitting network 0.3 dB - 3 to 3.5 dB total. When the antenna end measures 6 dB short in the field, eight times out of ten two of these items came in twice as bad as estimated. Rule 2: verify every specification by worst-case stacking, never typical values. 50 dB gain across four stages: front-end 18, driver 15, pre-final 10, final 7; each stage ±1 dB, worst-case sum ±4 dB. The ±1 dB flatness target is recovered by interstage equalizers - and the equalizer slope must be computed at both temperature extremes; set it right at hot and cold and the middle temperatures take care of themselves. Rule 3: thermal design with +20 °C ambient margin. Rated to 50 °C ambient? Compute junction at 70 °C. Junction red line 150 °C, design target under 125 °C - the remaining 25 °C absorb blocked airflow, dust, solar load and other field variables. 16. Five Case Studies, From Kickoff to DeliveryWhere the previous sections covered "which specification is the lifeline", this section lays five real projects completely open: how the customer stated the need, how the block diagram was fixed, which circuit trade-offs were made, where the prototype stuck, and how it was delivered. Numbers are from the measured records of the time; device models are indicative only - always re-verify datasheets and lot data before replicating. Customers are anonymized; only the use is named. Case 1: border-line fixed stations - a 100 W module running 24/7How the customer asked. A low-altitude security integrator deploying twenty-plus fixed stations along a border, each a transmit front end in a 3 m pole-mounted cabinet. The verbal requirement was one sentence: "8-12 GHz full band, 100 W, 24/7, and I can't climb up there when it fails." Translated: single-tone saturation ≥110 W, full-band flatness within ±1.5 dB, harmonics better than -50 dBc, survivable 5:1 load mismatch, -20 to +55 °C ambient with sunlit cabinet case temperatures potentially 75 °C, five-year life, failure rate under 0.3%. How the block diagram was fixed. Pure straight amplification, no conversion: input limiter plus digital attenuator (0-31 dB, 0.5 dB steps) managing level; two LNA-driver stages for front-end gain; an SPDT switch into a calibration path; final stage two-way quadrature combined. Why no conversion? The customer wanted broadband jamming with the signal source in their own baseband box - we only amplify; a conversion chain here is pure cost. But I left one switch and an external-LO interface after the driver stage - a habit born of scars: nine times out of ten, phase two of such projects adds "single-tone precision strike"; with the interface reserved, phase two swaps one board; without it, the whole unit is redone. Two key trade-offs. First, the final-stage device: domestic SiC-substrate GaN die, selected at 150 W-class saturation (datasheet pulsed; derated 50% for CW ≈120 W); thermal resistance 3.2 °C/W - the number that set the entire cooling solution. Combining via quadrature hybrid, two paths held within 0.3 dB amplitude and 5° phase, measured combining efficiency 90%. Second, the filter: the cabinet lives outdoors and out-of-band spurious had to pass regulator acceptance, so a cavity low-pass after the circulator plus an absorptive microstrip low-pass at the front-end transition; measured system harmonics -55 to -60 dBc, 5-10 dB of margin. Where the prototype stuck. The first prototype sailed through the 25 °C lab, then failed in the sun-exposure test shed: after four hours continuous, case at 68 °C, output down 2 dB; half an hour later over-temp tripped. The root cause was not insufficient heat sink but the thermal interface: AuSn die attach to MoCu carrier (fine), then carrier to heat sink through a thermal pad - 3 °C·cm²/W over only 6 cm², that single layer eating 0.5 °C/W. Replaced with indium foil plus screws at fixed torque (1.2 N·m each), same conditions dropped case temperature to 59 °C with no power sag. Nearly every newcomer steps on this: computing thermal resistance from die to heat sink while forgetting the "thin" layer in between. Two extras delivered. Software case-temperature derating (linear fold from 60 °C, 60% at 70 °C) written into the acceptance document - the customer accepted. And a "thermal health record" per unit: full-load steady-state case temperature, all temperature points, gate voltages, currents - field telemetry compared against baseline; case temperature 5 °C above baseline schedules maintenance. Twenty-plus stations ran three years; only two units returned - one fan, one sand-intruded waterproof connector. Case 2: vehicular platform - 100 W with vibration and supply as the two mountainsRequirement character. The vehicle works while moving; equipment in the equipment bay below the roof turntable; random vibration per the general vehicular profile, worst vertical. Supply: 24 V battery, dipping to 16 V at start, spiking above 40 V under load steps. Band still 8-12 GHz at 100 W. What differed from Case 1. Cavity: Case 1's bolted-plus-gasket assembly is unacceptable under vibration - gasket compression shifts and insertion loss drifts. This unit: machined-from-solid aluminum cavity, cover with twelve screws plus conductive silicone strip; 300 g heavier, post-vibration insertion-loss shift within 0.1 dB. Filter tuning screws all staked - tuned first, witness-marked, then staked with single-component silicone. Supply entrance: TVS + pi filter + ideal-diode reverse protection; on undervoltage, not shutdown but maximum attenuation while preserving final-stage bias, auto-recovery when voltage returns. This logic was rescued from the field: the early version rebooted on every jolt, and the customer waited two minutes per re-setup, furiously. The most time-consuming integration item. On-vehicle radiated tests showed some azimuths reading 3 dB below the bench. Two days of suspicion later, the culprit was the turntable cable: a 1.5 m flexible cable under bending changed both phase and loss - at one bend angle, 1.2 dB. Low-loss stabilizing cable plus a minimum 80 mm bend radius fixed it. The conclusion went into the assembly standard: every RF cable on the vehicle is retested for loss at maximum bend after installation. Case 3: airborne pod - 100 W-class narrowband jammer unit at weight-and-thermal double limitThe most "high power in a confined volume" case. Available pod volume: 200×120×40 mm; mass budget 1.8 kg; environment -40 to +70 °C plus low pressure. Band: one sub-band, 800 MHz wide, 80 W. The trade-offs. No room for a large heat sink, so the philosophy flipped from "conduct heat out" to "shorten the heat path": device soldered directly onto the thick-walled housing, the housing's outer surface itself the radiating surface, mated to a heat-pipe cold plate. Housing material: AlSi, CTE 7.5 ppm/K - close to SiC's 4 - giving solder-layer thermal-cycling life an order of magnitude better than soldering to aluminum. Matching networks entirely on co-fired ceramic substrates: half the footprint of microstrip, superb consistency - the precondition for daring "tuning-free" on a narrowband module. Filtering: one dielectric filter, 0.8 dB in-band loss, very steep skirts - at the price of temperature drift; so the dielectric filter carries a heater resistor and sensor, thermostatted at 55 °C. Two watts of extra consumption buys band-edge drift of only 2 MHz across temperature cycling. The cold-start pit. At -40 °C the gate threshold drifts; a fixed gate voltage at power-up slams current. Fix: DAC-set gate voltage - start conservative, read current, then step toward target quiescent current via a temperature lookup table, all within 200 ms. This sequence was later adopted on every product as the standard power-up profile. Case 4: counter-UAS radar T/R module - isolation and coherence, not powerRequirement. A low-altitude surveillance radar, one T/R channel per element: 100 W peak transmit, 10 µs pulse, 10% duty; receive NF better than 3 dB. The difficulty was never transmit power but T/R isolation: one antenna port, 100 W during transmit - the receive front end must survive. How it was solved. Three things. First, circulator plus two limiter stages: circulator reverse isolation 20 dB, then a PIN limiter plus a diode limiter pressing leakage into the receiver to milliwatts; limiter recovery time under 2 µs, or the receiver is blind for tens of microseconds after each pulse and near targets are lost. Second, shared antenna port switched by a PIN switch rated 200 W, 0.4 dB insertion loss - do not dismiss 0.4 dB: it is paid simultaneously from transmit power and receive NF. Third, coherence: one shared LO, distribution network all equal-length within 2 mm (≈3° of phase error). Final channel consistency: 0.6 dB amplitude, 5° phase - achieved not by tuning but by ceramic substrates plus automated assembly. The domestic-substitution episode. The receive LNA chain originally used an imported die; substitution selected a domestic GaAs E-mode LNA, NF 1.2 dB, gain 18 dB - specs matched, but the datasheet omitted large-signal damage threshold. In testing, a false pulse (100 W through 35 dB coupling) killed the third device: the internal first-gate had no protection. Adding a series limiter diode at the input raised the damage threshold to safety at 0.1 dB NF penalty - fully acceptable. Lesson: when substituting LNAs, NF and gain alone are not enough - damage threshold must be measured. Case 5: LEO payload front end - the full path from ground prototype to flight hardwareThis case best proves "the supplier matters more than the design". Mission: an LEO small-satellite RF payload front end, near X band, 500 MHz bandwidth, 80 W pulsed output, three-year orbital life. The design itself resembles Case 3: ceramic substrates, device soldered to housing, thermostatted dielectric filter. The real watershed was process and quality system. The path. Step one: ground prototype assembled from domestic modules to prove the link - two months. Step two: engineering models move to space-grade discipline: radiation data for every device (substitute or run your own TID test where absent); adhesives, wire insulation and potting through outgassing tests (TML and CVCM both passing); whisker-mitigated soldering. Step three: qualification units through the full suite - thermal vacuum, thermal cycling, random vibration, TID, single-event - with before/after performance deltas written into the report. Step four: flight and spare units with full traceability: device lot records, operator and inspection records per process step. Two memorable events. First, micro-assembly: the final-stage amplifier unit and driver chain were delivered as standard modules by a space-grade micro-assembly supplier, with interfaces aligned from day one - saving the die-bonding process ramp that, by our own experience, costs at least six months. Two rounds of device re-verification before launch were coordinated directly through that supplier's lot load-pull comparisons and radiation documentation. Second, a thermal-vacuum surprise: in the first TVAC cycle, output power dropped 0.8 dB at the cold plateau - the dielectric filter's thermostat heater lost its convection path in vacuum and the control loop's integral terms no longer fit; retuning the control parameters fixed it. Such problems never appear on the ground; only test exposes them. What the five cases teach together. Ground fixed stations test heat and longevity; vehicles test structure and supply; airborne tests volume and heat path; radar modules test isolation and consistency; space tests process and quality system. The hardware skeleton is always "GaN final stage + quadrature combining + filter protection", but what decides each project is the most easily ignored link in that scenario's chain: interface thermal resistance, cable bending, cold start, limiter recovery, outgassing. Find that link first, then talk specifications. 17. Every Specification, Derived BackwardsThe rookie treats the specification sheet as an acceptance checklist, testing line by line. The veteran works in reverse: from the outermost number inward, down to a single capacitor, a trace, a solder joint. Here the main specifications of this machine, derived one by one - afterwards, "design" stops being an abstraction. 17.1 From "100 W" to "how large a device"Outermost: port output 100 W (+50 dBm), typical 120 W, floor 90 W. One layer in: ahead of the port sits the output low-pass filter, 0.5-1.2 dB loss, take 0.8 - so the combining network must deliver +50.8 dBm ≈120 W. Next: combining network loss 0.4 dB (two-way quadrature) plus combining-efficiency loss 0.2 dB (amplitude/phase mismatch) - each device must reach +51.4 dBm ≈138 W. Next: the device's own output-match loss is already inside its load-pull result, but output DC block and test-coupler main-line loss add 0.2 dB - the die must deliver +51.6 dBm ≈145 W. Next: saturated power sags with temperature and lot - GaN saturation droops roughly 0.001-0.003 dB/°C; 125 °C of junction rise costs 0.12-0.4 dB, plus 1 dB for lot spread. Conclusion: select devices with saturation ≥160 W (+52 dBm), or two devices at 80-100 W each. That is where "100 W" lands in device selection. Pick a nominal-120 W device and the ledger balances exactly on paper - with zero real margin, and production will sink below 90 W. 17.2 From "±1 dB flatness" to "how much ripple each stage may have"System flatness ±1 dB is the stacked result of four or five gain stages, several matching networks, a filter and assorted traces. Engineering practice assigns ±0.2 dB per stage - four stages make ±0.8, leaving 0.2 dB for interconnects and connectors. Each stage's own requirement: no more than ±0.2 dB ripple across the octave. That sounds loose and is not - because a single stage's matching-network gain shape alone carries 0.3-0.5 dB of ripple, you must rely on interstage "complementation": the previous stage's top rises where the next stage's top falls, and the sum is flat. So interstage matching is never tuned stage-by-stage - it is optimized as one whole chain in simulation. The classic rookie error is tuning every stage "individually flattest" and ending with a worse system ripple. The filter's 0.2 dB is harsher: a cavity filter at ±0.3 dB passband ripple is routine; ±0.1 dB doubles machining and tuning cost. So budget the filter at 0.3 dB ripple and design the ALC compensation range to ±1 dB (digital attenuator plus VGA), letting the closed loop iron out the filter ripple - the filter handles only rejection, the loop owns flatness. Once that division is fixed, filter cost drops immediately and system flatness improves. A textbook "buy performance with circuits instead of machining" trade. 17.3 From "harmonics -50 dBc" to "how many filter poles"Harmonic -50 dBc, designed at -55 (5 dB process margin). The device's own second harmonic sits at -15 to -25 dBc depending on matching (with second-harmonic shorting, -30) - so the filter must add 25-40 dB. A low-pass rolls off about 6 dB per pole per octave (Butterworth) to 12 dB (Chebyshev). The second harmonic starts at 16 GHz, only 1.33× the passband edge - not even one octave of roll-off distance. Slope alone cannot do it; you need transmission zeros - one or two deep zeros placed at 16-18 GHz to push the harmonic down. This is what coupling probes and cross-coupling inside a cavity filter are for: not steeper slopes, but zeros exactly where you want them. Conclusion: a 7-9-pole cavity filter with one or two transmission zeros at 16.5 and 17.5 GHz (covering the low-end harmonics); above 18 GHz, rely on the cavity's parasitic rejection and housing absorption. The third harmonic (24-36 GHz) lies farther out and falls to ordinary roll-off. 17.4 From "5:1 load VSWR without damage" to "circulator plus protection"5:1 VSWR ⇒ reflection coefficient (5-1)/(5+1) = 0.67; reflected power = 0.67² ≈44% of incident. At 100 W output, worst case 44 W comes back. How long can a GaN device survive 44 W dumped into the die? The die's thermal capacity is tiny - microseconds push the junction past its limit. Without protection, one mismatch event is one dead device. The solution has two layers. Layer one, the circulator: reflected power is routed to the reverse-port absorption load, never back into the device. Rate the circulator at 3× nominal - a 100 W CW system needs a 300 W circulator, because under mismatch it passes 144 W. The reverse load also rated 300 W and rated for sustained dissipation (some mismatches persist). Layer two, protection: reverse-coupled detection - power folding from 10-15 W reflected, hard shutdown at 25 W, response in microseconds. Two layers together: the circulator blocks most, the electronics handle what leaks (including mismatches on either side of the circulator). A related subtlety: the circulator itself loses 0.6-1.2 dB and its isolation is finite (18-25 dB) - a few percent of reflection still gets through. That is why a circulator alone is never enough; electronic protection is mandatory. 17.5 From "-20 to +50 °C" to "what to test specially at the cold end"The cold end fails more often than the hot end - a fact newcomers miss. At low temperature: device gain rises (GaN's gain temperature coefficient is negative; -20 °C versus +25 °C gains 0.6-0.9 dB); gate thresholds drift; capacitance shifts (C0G barely moves; X7R can move 15%); battery and supply efficiency changes; TIMs harden and contact thermal resistance can grow; potting and adhesives shrink and can crack solder joints. So cold testing is not "does it work" but: with gain up, can the ALC still hold (does the attenuator still have downward range - why attenuation range needs bidirectional margin, not one-sided)? Is the detector's cold-end drift calibrated (calibration tables must cover -20 °C)? Does power-up into condensation short the board (pull the unit from the cold chamber, power it, and 28 V meets a wet board - power up inside the chamber, or conformal-coat and dry first)? 17.6 From "28 V, 8.1 A quiescent, 15.5 A at 50 W" to "supply and cabling"The least conspicuous numbers on the sheet, and the most revealing of design quality. Quiescent 8.1 A says the stage quiescents sum there: final ≈3 A, driver 1.5 A, front 1 A, pre-driver 0.5 A, plus control and bias overhead. At 50 W output, 15.5 A: the current increment is 7.4 A × 28 V = 207 W DC for 50 W RF - an "incremental efficiency" of 24%, plausible. At 120 W and 30% efficiency: 400 W DC in, 14.3 A, plus the quiescent base - total 17-18 A. That is why over-current protection sits at 20 A: 15% margin without nuisance trips. Supply design consequences: the 28 V bus sustains 20 A, transient 25 A+; connector power pins rated 10 A each, two in parallel; on-board 28 V traces at the rule-of-thumb 1 A per mm of width (1 oz copper, 20 °C rise) need 20 mm of copper for 20 A - or two layers paralleled with via stitching. My practice: open the solder mask over power traces and let reflow tin them, effectively doubling copper thickness. Cabling: a 2 m supply cable carrying 20 A with 0.5 V allowed drop needs 25 milliohms round trip - 4 m round trip, 6.25 milliohms per meter, about 2.7 mm² cross-section; select 4 mm². Undersized cable drags the module's input to 26 V and below, and power and efficiency fall with it - the customer blames the amplifier when the fault is the cable. I have watched this exact scene more than once. 18. Noise, Spurious and Phase Noise: Three Ledgers of In-Band CleanlinessPower, efficiency and flatness are the "big" specifications; noise, spurious and phase noise are the fine print - but the fine print decides whether the customer complains. 18.1 Broadband noise floor: an octave PA's built-in burdenA 100 W amplifier with 50 dB gain and, say, 10 dB noise figure (typical for broadband high-power PAs - interstage loss and late-stage noise contributions): thermal noise density -174 dBm/Hz, plus 10·log(4 GHz bandwidth) = 96 dB-Hz, plus NF, plus gain - the computed output noise density exceeds saturation, which tells you the truth: in a real amplifier the noise is clamped by saturation, not amplified forever. Practically, total output broadband noise sits 10-15 dB below saturated power - a few watts to ten watts for a 100 W PA, appearing on the analyzer as a raised noise pedestal across the whole band. Two consequences: in EMC-sensitive scenarios this pedestal itself is out-of-band interference; and it consumes dynamic range - signal power plus noise power must stay under saturation. Countermeasures: a band-pass filter before the driver stage removes out-of-band noise first (amplifying out-of-band noise is pure waste); low-NF first-stage devices (NF is set by the first stage - Friis, non-negotiable, so pay for the first stage); and gain allocation keeping first-stage gain high (20 dB or more) so later stages' noise contributions become negligible. 18.2 The spurious-source checklist, in search order
18.3 Phase noise: when it matters and how to manage itPhase noise matters in exactly two cases: the output is a narrowband tone or modulated signal (phase noise degrades BER and skirts into adjacent channels), or the machine serves radar/coherent systems (phase noise directly sets MTI clutter rejection). For an 8-12 GHz output at -90 dBc/Hz @10 kHz, the LO needs about -95 (the 20·log(f_LO/f_IF) correction is roughly 10 dB at a 6.4 GHz LO with 2 GHz IF). Getting there: DRHO or high-Q VCO as LO source; the highest practical PFD frequency (higher PFD widens the in-loop noise-suppression bandwidth); low-noise LDOs powering the VCO and PLL (switching-supply noise modulating the VCO easily costs 10 dB of phase noise). A measurement trick: a spectrum analyzer cannot measure phase noise directly - its own phase noise masks the signal. Either use the analyzer's phase-noise option (with cross-correlation), or the delay-line discriminator: split the signal, delay one path through a long line, mix, convert phase noise into voltage noise, measure with a low-noise amplifier and analyzer. Simple equipment, the most practical laboratory method. 19. Thermal Design: The Complete Chain, 300 Watts From Die to Air (or Water)19.1 The heat-source inventoryFinal stage: two devices, 60 W RF out each (120 W combined), 55% efficiency each → 109 W DC in, 49 W dissipated each, 98 W total. Driver stage: its output is the final stage's input (60 W - 10 dB gain = 6 W per device, 12 W total); at 40% efficiency, 30 W DC, 18 W dissipated. Front driver: 2 W out, 35% efficiency, 3.7 W dissipated. Front-end and conversion chain: about 10 W, all heat. Passive losses become heat: combining 0.4 dB (≈11 W), filter 0.8 dB (≈22 W), circulator 0.8 dB (≈22 W) - 55 W. Total dissipation: 98 + 18 + 3.7 + 10 + 55 ≈185 W; plus supply conversion loss (90% efficient, 40 W of 400 W in) - internal heat ≈225 W. Worst case (22% efficiency, 90 W output): 409 W DC in, 319 W dissipated, plus 35 W supply loss = 354 W. Design to 354 W. 19.2 The thermal-resistance chain, segment by segmentDie to case: 3 °C/W is a selection-stage number - at 49 W per device the junction-case delta would be 147 °C, which is unacceptable, and reveals the real constraint: die thermal resistance is quoted over the whole die area, and power density is the true limit. Engineering practice reads the datasheet's maximum dissipation and junction-temperature derating curves. So the final-stage device (160 W saturation class, metal-ceramic or flange package) carries a junction-case thermal resistance typically 1-1.5 °C/W: 49 W gives a 50-75 °C rise.
Junction-to-coolant total: 80-105 °C. Coolant inlet 35 °C (typical ground/vehicular) → junction 115-140 °C - inside GaN's 175 °C limit with a tier of margin. Air cooling instead: a heat sink at 0.2 °C/W (354 W, 70 °C rise) demands several hundred CFM plus a very large sink - it does not fit in 160×120×25 mm. Liquid cooling is the only solution - a conclusion computed, not assumed. 19.3 The three most-overlooked thermal itemsFirst, CTE matching. GaN-on-SiC ≈4 ppm/K, copper 17, aluminum 23, MoCu 6-7, CuW 6-8: a device soldered directly to aluminum cracks its solder layer within a few hundred thermal cycles - the MoCu/CuW transition carrier is money buying lifetime. Second, heat-flux concentration: 49 W leaving a few mm² of die approaches 10 W/mm² - at that level any void, gap or contamination in any interface becomes a hot spot. Hold eutectic/solder voids under 5% and sample-inspect by acoustic microscopy or X-ray; skip this and hundreds of units into production, "random device burnouts" erupt. Third, temperature measurement: the thermistor on the carrier (as above), plus in production a recorded full-load steady-state case temperature per unit as the "thermal health record" baseline - field case temperature 5 °C above baseline means aging interfaces or blocked airflow/water paths, enabling maintenance before failure. This data costs nothing and converts service from "fix when broken" to "replace when about to". 20. Component Selection: The Complete List and Substitution NotesThe whole machine, inside out, by position: what to look at, how to select domestically, what to beware when substituting. Models are public reference points only - always defer to current datasheets and measurements, and keep a second source for every position. 20.1 Final-stage GaN power devicesWhat to look at: saturated power (CW or pulsed - always ask), optimum load impedance versus frequency (load-pull data), gain (small-signal and large-signal), efficiency, junction-case thermal resistance, Vth distribution, package parasitics, operating voltage (28 V or 48 V). Domestic selection: CETC 13th/55th Institute ecosystem packaged devices and dies - the fullest reliability data, military channels with lifetime traceability and radiation tolerance, at the price of long lead times and slow sample requests; Guobo Electronics - the 55th Institute's industrialization platform, complete LF-to-Ku product line, strong production economics (military grade via the institute channel, volume via Guobo); Huatai Electronics - 0.1-6 GHz, 15 W drivers to 200 W final stages with ready-made two- and three-stage combinations for fast chain assembly; Huiruixin, Sichuan Yifeng, Junweixin, Chengchang, Xinbaite, Shidai Suxin - each with band/power/package strengths, some die, some packaged, some internally matched modules; Suzhou Nengxun, Haiwei Huaxin, Sanan Optoelectronics - foundry and device supplier both, strong for custom work. Substitution cautions: never swap devices, imported or domestic, on "similar power and frequency" alone. Redo four things every time: load-pull, matching simulation, large- and small-signal measurement, thermal testing. Package format (metal-ceramic, plastic, internally matched module, die) changes parasitics enormously - changing package equals changing the match. Vth spreads widely, so gate bias must be adjustable, and production calibration must include a "set quiescent current" step. 20.2 Drivers and front endsPlastic GaN small-signal power (10-15 W, 28 V, 0.1-6/7 GHz, 14-15 dB gain, >60% efficiency, DFN 4×4 or 5×5): the universal driver-stage part - tens of RMB, production-stable, one part covering many bands. GaAs pHEMT gain blocks (1-7 GHz, >30 dB gain, 10:1 VSWR survival): front ends and receive chains - cheap and rugged. Caution: GaAs breakdown is only 5-8 V - guard against overvoltage and ESD; series gate resistors mandatory, ESD-safe soldering stations mandatory. 20.3 LNAsWhat to look at: NF, gain, OIP3 and P1dB, in-band flatness, out-of-band stability. 8-12 GHz GaAs LNAs: NF 1.5-2.5 dB, gain 15-25 dB mainstream. The classic failure is compression - behind a broadband LNA, one strong in-band signal into a sub-0-dBm P1dB crushes the whole chain, symptom "sensitivity mysteriously down 10 dB". Require P1dB above +10 to +15 dBm and put a limiter in front. Domestic GaAs LNA supply is ample - Guobo, OCR and a cohort of MMIC specialists all carry broadband catalog parts. 20.4 Filters, circulators, couplers, switches: the custom categoriesAll four are custom-first: lead time 2-4 weeks, order early, and write power handling, insertion loss, directivity, rejection and temperature range into the technical agreement. Domestic ferrite and cavity vendors (CETC institutes plus private firms) are mature. Key points: cavity filter tolerance ±0.05 mm, roughness 1.6 µm, silver 5-10 µm; circulators rated 3× system power; coupler directivity above 15 dB (below that, load VSWR corrupts the forward reading and drags the ALC loop off); at the 100 W level, switching is mechanical-coaxial or ferrite only - solid-state switches live in small-signal positions only. 20.5 Conversion and frequency sourcesPLL and VCO: phase noise, PFD ceiling, spurious suppression, lock time. Domestic options: CETC institute PLL ICs and frequency-source specialists; integrated VCO+PLL modules exist in 8-12 GHz - buying the module saves time, at the price of accepting its measured phase noise and spurious. Mixers: conversion loss, LO-RF isolation, IP3, LO drive - double-balanced is mature in this band. DACs and DDS: SFDR first - its spurs land directly on the output spectrum and cannot be filtered; then sample rate and output bandwidth. 20.6 Digital and controlFPGA: Unisoc (Logos, Titan), Anlogic (Elf, Eagle), Gowin (Arang/Gowin series) - industrial grade, low cost, domestic tool chains; thousands to tens of thousands of LUTs suffice for control duty. High-reliability and space: Fudan Micro rad-tolerant series, with ECC/scrubbing on configuration storage and triple-mode redundancy on critical state machines. Selection criteria: logic resources, PLL/clock management, I/O standards (3.3 V and 1.8 V both, LVDS preferred), temperature grade. MCU: domestic Arm-core parts from GigaDevice, Huada Semiconductor, Nations, Geehy, Sinowealth all suffice. Signal chain: 12-bit sub-Msps ADC for detector/temperature/current; op-amps and comparators from 3Peak, Novosense, SGMicro, Runshi; digital isolators from Chuantu. When substituting imports, verify pin compatibility, electrical parameters and temperature grades, and re-derive the analog front end. 20.7 Supply and passivesBucks and sequencing: Silergy, SGMicro, Southchip, Jedonpower all carry suitable converters, load switches and sequencers; gate negative bias from a negative charge pump plus LDO. Capacitors: Fenghua Advanced, Sunlord-class and TDK-equivalent domestic high-frequency MLCCs (0402, C0G, 50 V); bulk storage in tantalum or polymer. Always check the SRF curve - a 0402 1 µF self-resonates in the tens of MHz and is useless for filtering 2 GHz noise. Inductors: Sunlord/Magtop wire-wound and film - two hard criteria: SRF above the working frequency, saturation current above 2× stage maximum. Resistors: Yageo-class thin-film (never thick-film at RF); sense resistors four-terminal. Laminate: Rogers 4350B/3000-class; domestic hydrocarbon laminates now approach 4350B at lower cost - verify permittivity lot consistency and loss tangent, and have the fabricator confirm actual stackup parameters. Connectors: RF from CNLiTech-class domestic makers; 100 W needs high-power versions (extended inner conductor, enlarged air gap) or 2.92 mm; DC via hybrid D-type with 10 A pins paralleled in pairs. 21. System-Level Integration: Counter-UAS Sets, Radars, Radios and Space PayloadsThe module passing its own tests is not the finish line - the real test is working inside a host system alongside other subsystems. This section is integration. 21.1 Counter-UAS equipment integrationA counter-UAS set typically contains: detection (passive DF or radar), C2, jamming transmit (where this PA lives), antenna positioner, power. Order and focus:
21.2 Radar RF modules and T/R integrationRadar scenarios revolve around "coherent" and "timing", entirely unlike jamming.
21.3 Ad-hoc network radio integrationMesh radios are broadband, frequency-hopping, multi-carrier - the PA faces high-PAR, wideband signals demanding linearity.
21.4 Space payload integrationSpace integration leaves evidence at every step.
22. The Test Instruments: Five Machines, and Why Every Knob Is Where It IsThis section explains settings and their failure modes - the operating steps are in the manuals. 22.1 Vector network analyzerCalibration: the unit works 8-12 GHz - cal kit rated past 12 GHz; electronic cal is easiest, mechanical cal means full two-port SOLT. Push the calibration plane onto the device pins: calibrating at the cable ends bakes cable and adapter phase into every S-parameter and the matching design is wrong throughout. Fix with port extension or fixture de-embedding. Power: small-signal at -10 to -20 dBm (keep the DUT linear). For large-signal S-parameters (compressed-state input/output impedance) use power-sweep mode, or switch to power meter plus tuner. IF bandwidth: 100 Hz-1 kHz - narrower is slow, wider is noisy. Measuring a deep filter stopband (below -60 dB) needs IFBW below 10 Hz, or the "stopband" you see is the noise floor, not real rejection. Classic errors: measuring without calibration; moving cables after calibration; no attenuation so the analyzer's receiver compresses (the receiver tolerates only about +10 dBm - a 100 W output needs 30-40 dB of fixed attenuation placed close to the DUT to limit cable radiation); misjudging VSWR from an unseated port connection. 22.2 Spectrum analyzerThree critical settings: RBW, VBW, input attenuation. RBW resolves close tones and sets the noise-floor reading: 100 kHz for spurious, 100 Hz-1 kHz for phase-noise skirts, 100 kHz-1 MHz for harmonics. VBW smooths the trace: typically 3× RBW; 10× RBW stabilizes noise-like readings; smaller for faint signals. Input attenuation is where front ends die: 100 W through 30 dB still leaves 100 mW (+20 dBm) - too hot for the analyzer (mixer compression), so 40-50 dB total, and with the preamp on, keep input below about -30 dBm. My habit: a 30 dB fixed attenuator permanently on the analyzer input, variable after it - hand tremor then cannot cost a front end. Correction: attenuator frequency response is not flat - above 10 GHz it adds several dB. Measure the attenuator's S21 across band (VNA) and load it into the analyzer's amplitude-correction table, or harmonic readings run 2-3 dB optimistic - you declare pass, the customer's bench says fail. 22.3 Power meterThe most accurate power instrument, three traps. First, sensor frequency response: thermal and diode sensors differ; choose one covering 8-12 GHz and zero/calibrate with the meter's reference source at every sensor or cable change (calibration at 50 MHz or 1 GHz; frequency corrections come from the sensor's internal cal-factor table). Second, average versus peak: CW on an average sensor; pulses need a peak sensor or triggering with correct video bandwidth and trigger level - or the "peak" is fiction. Pulse power = average / duty cycle, but both measured on the same instrument; mixing instruments costs several dB. Third, sensor survivability: 100 W (+50 dBm) far exceeds typical sensor ceilings (+20 to +30 dBm) - always attenuate or couple first, and measure the coupler's actual coupling at each frequency; never trust the nameplate. 22.4 OscilloscopeIn PA work the scope watches three things: pulse envelope, detector signals, supply ripple. Envelope: via an active RF probe (bandwidth above the carrier) or a detector converting RF to video; measure rise, fall, droop, overshoot, jitter - with the shortest possible ground connection (spring ground pin), or the ground lead's inductance resonating with probe capacitance stacks ringing onto the waveform and you chase a circuit fault that lives in the probe. Detector signals: watch ALC loop stability - an unstable loop shows a low-frequency oscillation (tens of Hz to kHz) on the detector voltage; faster than hunting sidebands on an analyzer. Supply ripple: AC coupling, 20 MHz bandwidth limit, ×1 probe with the ground clip removed (ground spring instead); on the drain, ripple must stay within tens of millivolts peak-peak. 22.5 Signal generatorThe generator is the "input", and its cleanliness decides whether the spurs you measure belong to the PA or to the generator. So: before measuring harmonics/spurious, connect the generator straight into the analyzer, record its own spur positions, and exclude them later - skip this and you spend days chasing a spur that never lived in the PA. Set output power accurately and attenuate - the PA input has limiter protection and an over-driven generator trips it or damages the front end. At large signals the generator itself compresses: measure actual drive at the PA input with a power meter; never trust the generator's display. Two-tone needs two generators plus a combiner (or one dual-channel generator), phase-noise-correlated via a shared reference clock, or measured IM3 is contaminated by phase noise; combiner isolation above 20 dB, or the two sources injection-lock each other and the result is fiction. 22.6 What I require on the bench before a unit goes upA 30 dB fixed attenuator (analyzer input), a 20 dB fixed attenuator (power meter/coupler), one short calibrated RF jumper, one manual variable attenuator (fine drive adjustment), an IR thermometer, a torque screwdriver (device mounting), an X-ray or borescope (joints), and a handwritten test log. With this laid out, measuring one unit takes half the time. 23. Fault Diagnosis: The Trees, Symptom by SymptomTwelve phenomena, each with the three most likely causes and the fastest verification:
24. Power Combining: The Chapter No 100 W Design EscapesA single device cannot reach 100 W saturation across 8-12 GHz; the standard route is two 50 W devices or four 30-40 W devices combined. The combining method decides how much power you finally get, how the flatness looks, and what happens when one path fails. 24.1 Choosing among three combining methodsWilkinson: best isolation (better than 20 dB between ports), easy bandwidth, ~0.3 dB loss. The weakness is power handling of the 100-ohm quarter-wave leg - at 10 GHz the microstrip is ~0.15 mm wide and 50 W per path will heat or arc it. If used, double copper to 70 µm or move to stripline/suspended microstrip. Branch-line hybrid (3 dB quadrature coupler): two-way combining with the isolated port terminated. A single section gives ~15% bandwidth - 8-12 GHz needs 3-4 cascaded sections, 0.4-0.6 dB loss. Strength: high tolerance to path phase error, and one failed path costs only 3 dB instead of everything, with limited VSWR degradation - the property vehicular and airborne platforms prize most. Balun / transmission-line transformer (coax, stripline): easiest to span an octave or more, loss down to 0.2 dB - but poor isolation: the two devices "see" each other, and one drifting device drags the other. Only for mature designs with well-controlled device consistency. 24.2 The combining-efficiency ledgerFour paths, 30 W target each, theoretical 120 W. Combining network 0.5 dB (≈11%) → 107 W; output isolator 0.3 dB (≈7%) → 99 W. That is one origin of the gap between "120 W typical" and "100 W minimum": typical assumes all four devices landed in the best lot; minimum must assume insertion loss at ceiling, device spread and temperature - all three worst-case simultaneously. Budget 0.8 dB for combining and 0.4 dB for the isolator - effectively 25% of device power removed on paper - so select devices at 35 W per path or better. 24.3 Amplitude/phase consistency: the invisible killer of flatnessTwo-way combining: ΔA = 0.5 dB costs about 0.2 dB; Δφ = 10° about 0.03 dB, 20° about 0.13 dB; ΔA = 1 dB with Δφ = 20° together costs about 0.4 dB - and it varies with frequency, directly eating 40% of a ±1 dB flatness budget. Control measures: all four paths equal-length from the input split point, length error within 0.5 mm (≈6° at 10 GHz); bends uniformly arc-shaped (radius ≥ 3× trace width) or chamfered, never right angles; identical via counts and layer transitions per path; tunable attenuators before the final stage (pi, 0.5 dB steps) and tunable phase (open stub plus screw, ±15°). In development, measure each path's S21 on a VNA, flatten amplitude and phase, then lock everything. 24.4 Do not economize on isolation loadsHybrid isolated ports and Wilkinson resistors all need loads. In normal operation they dissipate almost nothing - but when one path fails or detunes, they absorb nearly one path's output. Rate them at 1.5× one path's power. For chip loads compute power density: a 10×5 mm alumina chip on a well-cooled baseplate handles 20-30 W long-term; beyond that, flange-mount. I have watched a 0.02-RMB 0402 resistor standing in for a load: one path detuned, the resistor charred and lifted its copper with it. 25. Gain Flatness and Linearity: ±1 dB and -50 dBc Harmonics, Delivered by Two Compensation LayersDevice gain falls with frequency - across the 8-12 GHz octave, a bare device slopes 4-6 dB. Without hand-tuning capacitors one by one, compensation must be built into the architecture. 25.1 Frequency compensation in two layersLayer one, interstage equalizers: commonly a branch-line coupler with a frequency-dependent absorptive load, or a reflective equalizer (high-impedance line plus shunt resistor). The goal is not flattening 50 dB in one stroke but pressing each stage's slope within ±0.5 dB - four stages worst-case ±2 dB, with the later-stage equalizer absorbing the remainder. Layer two, input pre-emphasis: a slope-reversed network before the driver gives 10-12 GHz an extra 1-1.5 dB, pre-compensating the final stage's natural high-end roll-off. Note this compensates small-signal gain; at saturation the device is compressed and the added gain does not translate proportionally into power. So flatness must be examined on two curves - small signal and saturated output; the specification's ±1 dB normally refers to saturated output. Do not hand over the small-signal curve as compliance evidence. 25.2 An equalizer cell you can use directlyAbsorptive: a 3 dB branch-line coupler, through arm terminated in series R-C to ground (R 100-200 ohm, C 0.3-0.8 pF), coupled arm into a matched load. At low frequency the capacitive reactance is large and almost nothing is absorbed; as frequency rises, reactance falls and absorption grows - a positive slope. R sets the slope magnitude, C sets the corner, corner ≈1/(2πRC): R = 150 ohm, C = 0.5 pF puts the corner at 2.1 GHz, approximating a straight line across 8-12 GHz. Two cells in series deliver 2-3 dB of compensation at 0.5-0.8 dB of insertion-loss price. 25.3 Three gates for harmonics at -50 dBcGate one, the device itself: GaN second harmonic in saturation runs -15 to -25 dBc; it will never reach -50 alone. Gate two, harmonic terminations in the output match: a controlled short or open at the second harmonic (16-24 GHz) presses it to -30 to -35 dBc; the third (24-36 GHz) falls to -25 dBc or so on output capacitance and package parasitics alone. Gate three, the output low-pass or cavity filter: cutoff at 13-14 GHz, 15-20 dB rejection above 16 GHz. Three gates stacked, second harmonic below -50 dBc with margin. Critical reminder: harmonics are measured at saturated output, so the filter works at full power. Never pass judgment from a small-signal S21 sweep - at high power cavities can intermodulate and regenerate harmonics. I have measured a cavity at -60 dBc small-signal that reached only -48 dBc at full power - the cause was one poorly-contacted tuning screw forming a nonlinear junction. 25.4 The search order for -60 dBc spuriousThree classes, searched fastest in this order: (1) LO and clock leakage and their harmonics (conversion architectures) - shut down the baseband; if the spurs vanish, found; (2) sidebands from switching-supply ripple - run once on battery; if they drop away, it is the supply; (3) digital (FPGA/DAC) clock coupling into the RF chain - inspect the ground seam and via fencing between digital and RF boards; clock traces must have unbroken ground beneath. To measure below -60 dBc, freeze the analyzer settings first: input attenuation 10 dB (never 0 - mixer compression fabricates spurs), preamp on, RBW 1 kHz, VBW 10 Hz, sweep time manually 5 s or more, reference level just below overload. Settings not frozen, half the "spurs" you see were born inside the instrument. 26. Closing ThoughtsThis trade, in the end, comes down to two things: heat and impedance. Think them through and the rest is craft; fail to, and however beautiful the simulation, the outcome is luck. Do not fear measurement. The most common rookie delay is not daring to measure an unfinished tune - the measured curve is the most honest teacher, and the people who spend their first year longest in the lab always advance fastest. Write down every failure. PA faults, over a career, number maybe thirty or forty kinds; fill one notebook and you can run projects on your own. Domestic sourcing is no longer a compromise: GaN, GaAs, FPGA, data converters, passives, laminate, connectors - the chain now closes domestically, with several suppliers able to deliver at module or even subsystem level. Do the homework: request datasheets, full-temperature data, lot-consistency data and second sources - holding that evidence is what lets you promise a customer with a straight face. Shenzhen Huashi Digital Technology Co., Ltd. (深圳市华视数码科技有限公司) independently develops, manufactures and sells this 100W-class 8-12 GHz broadband RF power amplifier module and the matching driver chains, control and protection electronics, supporting custom filters, combining networks and system-level integration for ground, vehicular, man-packed and space-heritage platforms. Full-temperature test data, lot-consistency reports and qualification documentation are available on request. For selection, customization and quotation, contact: Sales Department 18688706870, website: www.szwisi.com. |