|
100 MHz-6.5 GHz Wideband Sweep Source Module (1/2): 200 MHz Bandwidth, 500 kHz Step, +5 dBmAt a glance
1. The difficulty is the octave count, not the powerA 65:1 frequency span is where this design lives or dies. At 100 MHz the free-space wavelength is 3 m; at 6.5 GHz it is 46 mm. A 20 mm microstrip line is an ideal wire at the bottom of the band and a transmission line approaching 90° of electrical length at the top. A 100 pF capacitor behaves as a capacitor at 100 MHz, passes self-resonance around 2 GHz, and is an ill-defined short stub by 6.5 GHz. Designers who apply narrowband habits to this problem get a predictable failure mode: gain stacked at the low end collapses at the high end; flatten the high end and the low end oscillates. Every adjustment trades one end of the band against the other. The technical content of this module therefore reduces to one sentence: hold gain, flatness, output power, spurious content, DC power and volume simultaneously across a 65:1 span. The sections below follow the order in which those constraints have to be resolved — requirement decomposition first, architecture second, component selection third, matching and production last.
2. Decomposing the requirement before drawing a schematic2.1 Four tiers, not one listThe most common early mistake is to treat every line of the specification table as equally binding. In practice the requirements fall into four tiers, and the engineering effort has to be allocated accordingly.
The value of the tiering shows up in design reviews: it makes visible which numbers can be traded and which cannot. The classic failure is spending tier-1 effort on a tier-4 target — doubling cost for a parameter nobody will measure. 2.2 Instantaneous or tunable? A three-fold cost differenceBefore choosing devices, establish whether 100–6500 MHz means instantaneous coverage or tunable coverage. The two answers differ by roughly a factor of three in cost. Instantaneous 6.4 GHz coverage forbids band segmentation in the RF path. It forces distributed amplifiers and broadband balun structures, and the noise figure, DC power and cost all suffer. Tunable coverage — only 200 MHz occupied at any instant — allows switched filter banks, band-split matching and band-split synthesis, which raises performance and lowers cost at the same time. Roughly ninety percent of sweep-source and interference-simulation requirements are the tunable case: settable centre frequency, configurable bandwidth, 200 MHz maximum. Once that is settled the architecture follows: the RF chain is a single 100–6500 MHz through-path, and frequency selection is done in the LO and filter banks, not in the RF path. 2.3 The 500 kHz figure: do the arithmetic first"500 kHz scan rate" is ambiguous. It normally means a 500 kHz step with configurable dwell, not 500,000 random-access hops per second. The arithmetic is worth doing explicitly. At 500 kHz steps across 100–6500 MHz there are 12,800 points. With 1 µs dwell a full sweep takes 12.8 ms; at 20 µs dwell, 0.26 s. Trivial in principle. The bottleneck is not PLL settling. It is the command interface. RS-422 at 115.2 kbps carries a set-frequency command of a dozen bytes in about 1 ms. If the host sends one command per point, the ceiling is roughly 1,000 hops per second — two and a half orders of magnitude short. The architecture therefore has to be table-driven: the host writes a hop table, dwell time and loop mode into FPGA RAM once; the FPGA then sequences the LO, attenuator and filter channel locally. That imposes two hard requirements on the synthesiser: PLL lock within 20 µs (10 µs preferred), and spurious settling that is acceptable within a 1 µs dwell. The digital side follows: 200 MHz instantaneous bandwidth needs at least 400 MSPS by Nyquist, so a 1 GSPS DAC with oversampling headroom. The DAC sample rate sets instantaneous bandwidth, not carrier frequency; translation to 100–6500 MHz is done by the upconversion LO. A 48-bit phase accumulator gives a frequency control word of about 1.41 × 10⁸ at 1 GSPS for a 500 kHz step, i.e. microhertz resolution — no post-tuning required. 2.4 The "+5 dBm" trap+5 dBm looks like a single-stage amplifier job. Three budgets show why it is not. Loss budget. A realistic accounting of the passive path: 0.2 dB per SMA (two), 0.3 dB for input limiting/ESD structure, 1.5–2.5 dB for the 6-bit DSA, 1.2 dB for two series switches, 2.5 dB for the worst filter segment, 0.3 dB for the coupler, 0.8 dB for microstrip and vias. Total 7–8 dB. If the final stage delivers +5 dBm, the port delivers −2 to −3 dBm. Margin budget. The specification says "greater than +5 dBm, may vary with frequency". That sentence is both an allowance and a trap: acceptance is done point by point with a power meter. The internal control line should be +8 dBm at any frequency, in any 200 MHz sub-band, over −40 to +70 °C, at 12 V minimum input — a 3 dB margin, which is the normal RF convention. Back-off budget. The final stage cannot run in compression; broadband noise modulation with a high peak-to-average ratio will clip and generate spurious products. With 6–8 dB of back-off, the final stage needs a P1dB of at least +18 to +20 dBm. Adding 7–8 dB of path loss, 8 dB of required output and 7 dB of back-off puts the target P1dB at +22 to +23 dBm. The three budgets converge on a concrete choice: a broadband driver with P1dB of about +23 dBm (200 mW) as the final stage, preceded by a ~15 dB small-signal gain block, total gain around 30 dB, driven at roughly 0 dBm. That is the cheapest configuration that meets the number. 2.5 31.5 dB in 0.5 dB steps: selecting the DSA0.5 dB steps over 31.5 dB is the standard 6-bit arrangement (0.5/1/2/4/8/16). Four criteria decide the part.
Two Huashi parts illustrate the trade:
Note the loss slope of Part A: 0.5 dB at the bottom, 1.3 dB at the top — already 0.8 dB of negative slope that must be carried in the flatness budget. Part B buys bandwidth and switching speed with 4 dB more loss; whether that is acceptable depends on whether the constraint is bandwidth or power budget. 2.6 Thermal: from module power to junction temperature12 V × 0.5 A = 6 W. That is the ceiling for the whole module at full load, and it is tighter than it looks. A realistic breakdown: RF chain quiescent 1.5–2 W (final stage 12 V × 130 mA ≈ 1.6 W, pre-drivers 0.3 W), FPGA and configuration 0.8–1.2 W, DAC and clock 0.5–0.8 W, RS-422, bias and detection 0.2 W. Total 3–4.5 W — inside 6 W, but without much room. The trap is FPGA logic: a large device with a fast clock can pass 2 W on its own. A ~50k-LUT class device is sufficient and keeps both power and thermal design simple. For spaceborne use the calculation has to go one level deeper. There is no convection in vacuum; heat leaves by conduction and radiation only. With a 25 °C/W junction-to-housing path, 1.6 W means a 40 °C rise; at a 70 °C housing limit the junction approaches 110 °C. GaAs pHEMT channels are rated to 150 °C, with long-term reliability normally derated to 120 °C. The final stage therefore needs a metal carrier (MoCu or CuMoCu) and a thickened base or an external thermal face. Carrying the budget through to junction temperature, rather than stopping at module power, is where the real margin is found. 2.7 The 220 g window6061 aluminium at 2.7 g/cm³: a 100 × 62 × 29 mm housing with 1.5 mm walls and a 2 mm base is about 55 g. Three boards with components, about 80 g. Four SMA connectors at 8 g each, 32 g. DC and I/O connectors, 20 g. Shielding walls and standoffs, 20 g. That is just over 200 g, leaving roughly 20 g for potting, thermal pads and harness. The rules that follow: do not thicken the walls; do not pot the whole cavity — pot only the PA and power sections; use lightweight stainless or beryllium-copper SMA bodies; machine the internal walls in one operation rather than assembling them from screws. Going further for space means aluminium-silicon or SiC carriers and pocketing — a different cost class. 3. Architecture3.1 Five master schematicsStart with five: RF chain, power, bias and protection, monitoring and communication, digital/FPGA. Sub-sheets are named by function, location and revision (e.g. 3.2 Transmit chain orderSMA in → ESD/limiter → 6-bit DSA → switch → band-split pre-filter →gain block (15 dB) → inter-stage equaliser → driver → final PA →directional coupler → output filter → isolator/limiter → SMA out The order is deliberate:
3.3 Coupler and detectorA −20 dB coupler feeds a log detector for power monitoring and closed-loop levelling. Coupling too tightly breaks the detector's dynamic range: at −10 dB coupling and +8 dBm output the coupled port sees −2 dBm, fine; but at maximum attenuation the output falls to −23 dBm and the coupled port to −33 dBm, outside a 50 dB detector's usable span. At −20 dB coupling the coupled port sees −12 dBm at full power (mid-scale, where the log slope is most linear) and −43 dBm at minimum output — still inside 50 dB of range. 3.4 Why not a single conversion to 6.5 GHzSingle conversion is the obvious approach: 200 MHz IF, LO swept 100–6700 MHz, mix straight to RF. Three obstacles make it unattractive.
The standard solution is dual conversion:
LO2 still needs two segments — 1.0–3.7 GHz and 3.7–7.4 GHz — with a divide-by-two or VCO switch at the crossover. Huashi fractional-N synthesisers in the 0.2–24 GHz class cover this with margin: −233 dBc/Hz normalised in integer mode, −228 dBc/Hz in fractional mode, 200 MHz integer / 150 MHz fractional phase-detector frequency, 3.3 V, 4 × 4 mm QFN. A phase-detector frequency above 100 MHz supports lock times under 20 µs. A caution on second-sourcing: substituting one synthesiser for another is not a package swap. Phase-noise curve, spur table, temperature drift and supply evidence have to be re-qualified. Any of the four missing means the substitution is not done. 3.5 Level plan, stage 2 upconversion
This is the single-tone, full-scale condition. During actual sweeping the DSA removes 10–20 dB, the final stage input falls to −4…+6 dBm, back-off exceeds 10 dB and the output settles at +8 to +10 dBm — on the internal control line. 3.6 Spurious budgetEvery mechanism needs an explicit line in the budget, otherwise the prototype spectrum is a field of grass with no way to trace it.
|