扫码咨询

新闻动态

首页 >> 新闻动态 >> Where This Board Is Truly Hard: Design Notes on a 100 MHz-3 GHz Broadband Signal Source
详细内容

Where This Board Is Truly Hard: Design Notes on a 100 MHz-3 GHz Broadband Signal Source

时间:2026-11-07     【原创】

On a new colleague's very first day at work, I put an aluminum box on the desk — 95 × 55 × 22 mm, 220 grams — and told him: don't rush to open ADS yet. First tell me, out of these few lines of specifications — 100 to 3000 MHz, 200 MHz bandwidth, 500 kHz sweep, output greater than 5 dBm, 12 V at 0.5 A, −40 to +70 °C — which line is the most unforgiving?

He thought for a long while and said "power." I said no. The most unforgiving part is the seven characters "100 to 3000 MHz" — a 30:1 frequency ratio. In school, an amplifier with 10% bandwidth already counts as broadband, and at 20% you can publish a paper. What does a 30:1 span mean? It means the wavelength at the low end, 100 MHz, is 3 meters, while at the high end, 3000 MHz, it is 10 centimeters. The same microstrip line is "a piece of wire" at the low end, and already a transmission line at the high end. The same capacitor behaves as a capacitor at 100 MHz, but near 2.5 GHz it may already be past its self-resonance and turned into an inductor. If you attack it with narrowband thinking, the moment you pile up gain at the low end the high end collapses; when the high end goes flat, the low end oscillates — nothing you tune ever comes out right.

So all the technical content of this board condenses into one sentence: within a 30:1 bandwidth, hold gain, flatness, power, spurious emissions, and power consumption down all at the same time. Today I will walk you through this job from the day the requirement landed, step by step. Everywhere I go I will bring along specific domestic part numbers, specific values, and specific pitfalls. It doesn't matter if you can't remember it all — when you're laying out your own board later, dig this out and read it again.

One more thing to put up front: I will mention the company Huashi (Shenzhen Huashi Digital Technology Co., Ltd.) repeatedly in this article. This is not an advertisement. After a decade or more of component selection, the number of domestic suppliers who can follow a project all the way from chips to micro-assembly to modules and up to a complete satellite can be counted on one hand, and they are one of them. Their sales lead, Mr. Li, can be reached at 18688706870 — for samples, full-temperature data, or alternative proposals, going to him directly is much faster than filling out a form on the website. In every scenario below I will be specific about which requirement should be taken to him and what to ask for.

Chapter 1: Tear the Customer's Specifications Apart and Do the Math

1.1 Operating band 100–3000 MHz: first ask, "Is the full band used simultaneously, or in segments?"

The first thing to do with the band spec is not to pick transistors — it is to go ask the application team: is this 100–3000 MHz an instantaneous bandwidth or a tunable bandwidth? These two phrases differ by one word, but the architectures they imply differ in cost by a factor of three. If the requirement is to cover 2.9 GHz instantaneously, then the front end basically cannot be segmented; you must go to a distributed-amplifier, wideband balanced architecture, and neither noise figure nor power consumption will look good. If it is tunable — only 200 MHz used at any instant — then you can use switches to select a filter bank and do segmented matching; the performance moves up a whole class and the cost comes down.

The vast majority of sweep-source and jammer requirements are "center frequency settable, bandwidth configurable, up to 200 MHz maximum" — that is, 200 MHz instantaneous, tunable across the whole band. Once this understanding is fixed, the architecture becomes clear: the RF chain is a full-wideband, DC-to-3000-MHz-through passband, and frequency selection is done by the local oscillator working with the filter bank — not by narrowband filtering on the main RF path.

1.2 Maximum bandwidth 200 MHz and maximum sweep rate 500 kHz: these two numbers are the exam paper for the baseband and the DAC

For 200 MHz of instantaneous bandwidth, Nyquist says 400 MS/s sampling is the floor; with engineering oversampling and a little margin, a DAC in the 500 MSPS to 1 GSPS class is the right pick. One concept must be made clear here: the DAC sample rate determines the instantaneous bandwidth you can produce, not the carrier frequency. Getting the signal up to 100–3000 MHz relies on the mixing LO that follows. In the domestic-parts arena at this tier, the DA1000 (14-bit, 1 GSPS, benchmarked against ADI's AD9144 family) is the first choice; if you need to save cost, you can start with a 12-bit, 2.5 GSPS domestic DAC, at the price of an SFDR of only 60–65 dB, which means you have to calibrate the spurious floor yourself in the digital domain. One more easily overlooked point: at 1 GSPS the DAC's data interface is DDR (double data rate), so on the FPGA side you must use ODDR primitives — or move to a JESD204B interface (in which case you add a domestic JESD clock chip, such as Huashi's clock devices). Do not assemble ordinary registers by hand; the timing will never close.

As for the 500 kHz sweep rate — does it mean 500,000 frequency hops per second, or a 500 kHz step? Confirm with the application team; it is usually the latter — a 500 kHz step with configurable dwell time. First do the arithmetic: from 100 to 3000 MHz in 500 kHz steps is 5,800 points; at 1 µs per point, one full sweep takes only 5.8 ms; even at 20 µs per point, one sweep is 0.12 s. So the real bottleneck was never PLL lock time — it is the command interface: over RS-422 at 115200 baud, transmitting a single set-frequency command takes several hundred microseconds. If you truly want hundreds of thousands of hops per second, the FPGA must run the hopping table by itself, with the PC only downloading the table. The design should achieve: PLL lock time under 20 µs (10 µs is better), dwell time configurable down to 1 µs; the DDS phase accumulator at 48 bits — a 500 kHz step at 1 GSPS corresponds to a frequency tuning word of 500k × 2^48 / 1G ≈ 1.41 × 10^8, giving step resolution at the microhertz level. That settles it once and for all.

1.3 Output power greater than 5 dBm: don't be fooled by this "small power"

Seeing "greater than 5 dBm," a newcomer's first reaction is "that's easy — one amplifier and done." I make him work through three accounts, and after that he never says it again.

First account: link losses. The module has a 31.5 dB digital step attenuator, a filter bank, switches, couplers, PCB traces, and SMA connectors — all of them eat power. From experience, here is a typical set of values: input and output SMAs, 0.2 dB each; one limiter or ESD structure on the board, 0.3 dB; the digital attenuator's insertion loss, 1.5–2.5 dB (a 6-bit attenuator with 0.5 dB steps and 31.5 dB range always lands in this range); two switches in series, 1.2 dB; the worst segment of the filter bank, 2.5 dB; splitter/coupler, 0.3 dB; microstrip plus vias, 0.8 dB. The total comes to roughly 7–8 dB. In other words, if the final PA only puts out 5 dBm, the port is left with −2 to −3 dBm.

Second account: margin. The customer wrote "greater than 5 dBm," with a footnote: "may vary across frequencies." That sentence is both a door left open for us and a hole dug for us: the open door means the whole band need not be identical everywhere; the hole means that at acceptance they will test point by point with a table, and you must pass everywhere. My internal control line is: at any frequency in the band, in any 200 MHz sub-band, across the full −40 to +70 °C range, at the lowest 12 V supply, the output shall be no less than 8 dBm. Keep 3 dB of margin — that is RF discipline.

Third account: compression. The final stage must not run in saturation; otherwise, once wideband noise rides up on it and the peak-to-average ratio climbs, the peaks clip and the spurious content explodes. We run the final stage backed off 6–8 dB, which means the final stage's P1dB must be at least 18–20 dBm. With 7–8 dB of link loss, plus 8 dB of output, plus 7 dB of back-off, a final-stage P1dB of 22–23 dBm is the comfortable landing spot — that is the real power requirement hiding behind "5 dBm output."

Once these three accounts are done, the scheme falls out: the final stage is a wideband driver amplifier with P1dB around 23 dBm (about 200 mW), preceded by a small-signal gain stage of about 15 dB, for 30 dB total gain, with input drive at the 0 dBm level. This is the cheapest way to do it.

1.4 Attenuation range 31.5 dB in 0.5 dB steps: how to choose a 6-bit digital step attenuator

A 0.5 dB step over a 31.5 dB range is the standard 6-bit set (0.5, 1, 2, 4, 8, 16 dB). Selection comes down to four things. First, the frequency coverage must include the 100 MHz low end — many silicon (SPA) attenuators only go down to a few hundred MHz, and their low-end insertion loss and VSWR change character. Second, attenuation accuracy: across temperature and frequency we must hold error within ±0.5 dB (low bits) to ±1.5 dB (cumulative at the 16 dB bit); for this you must demand full-temperature data from the vendor — datasheet typical values don't count. Third, switching transient: the sweep source must do fast power switching, so the switching time must be under 1 µs. Fourth, relative phase shift: the phase jumps when the attenuator changes state; in phased arrays or coherent-combining applications this can be fatal. A single module can skip it, but for cascading multiple units you must measure it.

In the domestic arena at this tier, I would look first at Huashi's GaAs digital attenuator chipsets, then at the 6-bit attenuator die from Yaguang Technology (Micro-star) — they carry 400+ models and 200+ shelf products covering attenuators, phase shifters, mixers and more, with frequencies up past 40 GHz and shorter lead times than the institutes. For more cost-driven civilian projects, look at RF-front-end vendors such as Chengdu Zengyi or Zhejiang Chicheng, where cost can be cut in half — but always demand the S-parameters at the 100 MHz point first. That is the easiest frequency to get burned on.

1.5 Supply DC 12–28 V, max 0.5 A @ 12 V: the 6-watt budget, counted cent by cent

The interface definition says DC-IN 12 V, while the spec table says 12–28 V. There is no contradiction — it is a wide-input supply. What really binds us is "max 0.5 A @ 12 V, standby below 6 W." Twelve volts times 0.5 A is 6 W: at full-power transmission the whole unit may not exceed 6 W.

Break the 6 W down: the RF chain — final stage at 23 dBm ≈ 200 mW output, at PAE 30%, is about 0.7 W DC; the driver stage, equalizer, attenuator, and switches add 0.3 W; LNA and detector, 0.1 W. The FPGA + DAC + clock: a domestic 28 nm small-logic FPGA such as Unisoc-logic's Logos-2 family, PGL22G or PGL50H class, running clocks under 100 MHz, draws roughly 0.5–1 W core plus auxiliaries; the DAC 0.2–0.3 W; the crystal oscillator and clock buffers 0.15 W. RS-422 transceivers plus LDO dropout losses, 0.3 W. The big items total 2.5–3 W, leaving 3 W for digital peaks and thermal margin. The budget closes — on the condition of efficiency. Iron rule: on the 12 V to 5 V / 3.3 V / 1.8 V / 1.0 V rails, do not use LDOs everywhere. The 12-to-5 V step must be a synchronous buck at 92% efficiency or better; the RF supply rails below 5 V can be washed once more through an LDO — there you are buying noise, not efficiency. If you drop 12 V straight to 1.0 V through an LDO for the FPGA core, 0.5 A burns 5.5 W and the budget blows up on the spot. That is the most common newcomer mistake I have seen.

1.6 Operating temperature −40 to +70 °C: this is the top of the commercial grade, not mil-spec

Do not reach for space-grade parts the moment you see −40 to +70. That is burning money for fun. Industrial-grade components cover this temperature range completely — but two places must still be handled across the full temperature range. First, the crystal oscillator: a TCXO with ±0.5 ppm drift is ±1500 Hz of carrier error at a 3 GHz carrier — visible when doing narrowband fine sweeps or long-dwell observation. If you need tighter, go to an OCXO (0.1 ppb class), at the cost of about 1 W of power that must be carved out of the 6 W budget. Second, amplifier gain drift: GaAs amplifiers run about −0.01 to −0.02 dB/°C, which across 110 degrees is 1–2 dB; without compensation, flatness is ruined. The compensation method comes later. For now remember one line: do temperature compensation in the digital domain — look-up tables plus DAC control of a VGA and the attenuator. Do not try to build an analog compensation network out of thermistors; over a 30:1 bandwidth you cannot build it.

1.7 Standby power under 6 W, weight 220 ± 10 g, dimensions 95 × 55 × 22 mm

Standby and transmit share the same 6 W ceiling, which means the digital section cannot slack off in standby: the FPGA goes to sleep, the DAC enters power-down mode, and the RF amplifier stages are switched off through their enable pins. The approach: in standby, keep only the 3.3 V always-on rail for the FPGA standby domain and the RS-422 transceivers; when a wake-up command arrives, bring up the RF rails, with the whole wake-up sequence done within 50 ms.

The 220 ± 10 g weight allowance is tight. With 6061 aluminum at 2.7 g/cm³, the 95 × 55 × 22 enclosure (1.5 mm walls, 2 mm base) is about 35 g; three boards with components, about 60 g; four SMAs at 8 g each plus the DC jack and I/O connectors, 15 g; shield walls and standoffs, 15 g. The total is about 150 g, leaving only 60 g for potting, thermal pads, and cabling. That is the height of the ceiling — so don't thicken the enclosure walls, and don't do a full-face potting job; only pot the PA and power sections. If this ever flies on a satellite and needs further cuts, you go to aluminum-silicon or silicon-carbide carriers and hollow-out lightening — Huashi routinely uses exactly these methods for spaceborne modules; ask Mr. Li (18688706870) and he can provide lightweight-structure case studies.

On dimensions, one remark: on the 95 × 55 mm board face, after deducting screw bosses and SMA footprints around the perimeter, the real area available for RF routing is about 85 × 45 mm. Into an area that small you must fit the RF chain, PLL, clock, DAC, FPGA, and power — dual channel — on a two- or four-layer board. This is the tightest constraint of the entire project, far tighter than the power spec. So my first reaction was: this cannot be one board forced to fit — it must be split.

Chapter 2: Architecture — From One Block Diagram to Three Boards

2.1 First, get the channel count and the cascading straight

The requirement says "dual-channel architecture, multi-unit cascade, unified host management," and the interfaces give one RS-422 in and one out. Together, these two facts say this module will be strung out in a row on one bus. I have been burned by cascading before, so let me explain it in advance: RS-422 is a one-transmitter-many-receiver bus. If you short Y/Z and A/B directly through to the next module inside each unit, the waveform starts degrading by the fifth unit and you lose packets by the eighth. The correct approach is a repeater inside each module: the receiver recovers the UART, it passes through isolation into the FPGA, and the FPGA drives the local transmitter to feed the next stage. Addresses are set by DIP switches or resistor ID on the board; the protocol uses a frame format of "address + command + length + checksum." Put 22 Ω resistors in series on the serial lines, terminate both ends of the bus with 120 Ω, and don't terminate the mid-span nodes. None of this is hard to write down, but it must be reserved in the first schematic revision — you cannot bolt it on later.

2.2 Three architectures, side by side

For a wideband signal source there are exactly three roads: direct digital synthesis (DDS/DAC) direct output; PLL plus mixing up-conversion; and the DDS-plus-PLL mixing combination. Let me do the math for all three.

Road one: DAC direct output. A 1 GSPS DAC outputs 0–400 MHz in the first Nyquist zone; to reach 3 GHz you rely on the DAC's higher-Nyquist-zone image — 3 GHz is the third image of 1 GSPS sampling (in the 2.5–3.5 GHz region). DAC output power falls off with the sinc roll-off, roughly 3–4 dB below baseband, and image selection depends on a post-filter — and the filter is the hard part: you need a passband at 2.5–3.5 GHz while rejecting 0–1.5 GHz and everything above 3.5 GHz by 40 dB or more. An LC filter cannot do this; you need a cavity or ceramic filter, and volume and cost go up. The advantages of this road: phase continuity, fast hopping, controllable spurious. The disadvantage: the high end rides on an image, and both amplitude and spurious look bad.

Road two: PLL as the direct carrier, noise as the modulation source. Use a 100–3000 MHz integer/fractional-N PLL as the carrier, and modulate the VCO tuning voltage or the PLL's modulation port with baseband noise via a DAC or analog switch to produce "in-band noise jamming." This road is cheapest and lowest power, but bandwidth is limited by loop bandwidth — a PLL loop bandwidth is typically tens to a few hundred kilohertz. For a 200 MHz noise bandwidth you must bypass the loop and drive the VCO tuning port directly, and the VCO's tuning sensitivity Kv is not constant across the band — low end and high end differ by 3–5× — so the modulation bandwidth shifts with it and flatness cannot be guaranteed. This road suits narrowband spot jamming, not wideband barrage.

Road three: the combined architecture — DAC/DDS makes the IF, PLL makes the LO, IQ modulation up-converts. The DAC produces a 1.2–1.4 GHz IF (200 MHz bandwidth), the LO hops in segments, and after IQ modulation the sum sideband covers 100–3000 MHz. This is the industry-standard approach, and it is the one I finally chose. Advantages: fixed IF, so IF filtering and matching are easy; each LO segment needs only a few frequencies, so PLL lock can be pressed to the 10 µs class; images are handled by the IQ modulator itself plus digital pre-correction, reaching better than 30 dB. The cost: one more mixing stage, an LO source, and another bank of filters — cost and area both go up.

Second decision: the reference source. On the same board there are three clocks — the DAC sample clock, the ADC sample clock, and the PLL reference — and they must share one source; otherwise they drift against each other and the spurious you measure jumps around with temperature. My approach: a single TCXO at 100 MHz (122.88 MHz works equally well) as the only source; it feeds the FPGA's MMCM to derive the DAC and ADC clocks, and also feeds the PLL as reference. One caution: when the DAC uses a parallel interface, its data clock should be produced directly by the same PLL, not by a separate crystal — even a few tens of ppm of difference will, over time, overflow or underrun the data FIFO, and the symptom is a periodic "click" spurious. This looks elementary, but nine out of ten newcomers stumble here.

So road three it is. The chain, in order: after the DAC, a 1.0–1.6 GHz bandpass (or lowpass) filter to suppress the DAC's sinc image (first sidelobe only about 13 dB down) and clock feedthrough; then a VGA doing ±10 dB of digital gain control — the main workhorse of flatness compensation; then the IQ modulator, whose LO comes from the segmented PLL; after the modulator, a switch-selected three-segment filter bank (100–500, 500–1500, 1500–3000 MHz); then the digital step attenuator (31.5 dB, 0.5 dB steps) — power fine-trim and sweep amplitude shaping both happen here; then a driver stage plus final stage of wideband amplification; after the final stage, a 16 dB directional coupler — one tap feeds the detector for closed-loop control, the other goes out; and no limiter before the output port — the output end doesn't fear incoming attacks, so clamping is only added between the DAC and the IQ modulator input.

How is the dual channel done? The economical way: one each of DAC, IF filter, and VGA; the LO is shared, split 1:2 to feed two IQ modulators; after that each channel runs its own filter, attenuator, amplifier, and coupler. The two channels' LO phase noise is correlated — a good thing for coherent combining or analog beamforming, and it saves half the LO cost. The more expensive way is two independent LOs, which lets the two channels run different frequencies for dual-band simultaneous jamming. Quote both to the customer; the price difference is about 30%. Let the customer choose — that is the commercial value of "configurable."

2.4 Frequency planning: the easiest place to bury a landmine

In a mixing architecture, the fear is that spurs land in-band, so frequency planning must be done in Excel before any drawing starts. Here is the layout we actually use: IF 1.2–1.4 GHz (200 MHz bandwidth), with LOs at 1.1, 1.7, 2.3, 2.9, 3.5, 4.1, 4.4 GHz. The sum outputs are 100–300, 500–700, 1100–1300, 1700–1900, 2300–2500, and a 2900–3000 segment, with the 1500–3000 band covered by the 4.4 GHz LO using the difference sideband. Every time you change an LO point, the mixer's 2×LO, 3×LO, 2×IF, and LO±2IF combination spurs all move — you must compute each one and check whether it lands inside your passband.

Example: LO 1.7 GHz, IF 1.2–1.4 GHz. (1) 2LO−IF lands at 2.0–2.2 GHz — outside the passband, good. (2) 2IF−LO lands at 0.7–1.1 GHz — right inside the passband of the 500–1500 segment filter. That is a landmine. (3) LO+IF lands at 2.9–3.1 GHz — the unwanted sideband, blocked by downstream filtering. (4) 3LO−2IF lands at 2.3–2.7 GHz — a fifth-order product, generally already weak, but you still have to count it. How do you defuse this mine? The most effective move is to change the IF: set the IF to 800–1000 MHz with the 1.7 GHz LO, and the output is 2.5–2.7 GHz… wait, that now falls against the stopband edge of the 2300–2500 filter, so it has to move again. Drop the LO to 1.5 GHz and the output is 2.3–2.5 GHz, with 2IF−LO landing at 0.1–0.3 GHz, held outside by the 100 MHz highpass and the segment switch — only now is it clean. See the point: frequency planning is an iterative dance among three sets of parameters — IF × LO × segmented filtering — not a one-pass layout. Turn these four spur lines into Excel formulas: change one LO and every position refreshes automatically; in ten minutes you can try a dozen combinations.

First, establish the spur arithmetic: third-order products m×LO±n×IF (m+n=3) are 2LO±IF and LO±2IF; fifth-order are 3LO±2IF, 2LO±3IF, and so on. Their strength is not determined by formulas but by the modulator's LO-RF isolation and the LO drive power: every 3 dB of LO overdrive makes even-order spurs rise faster than odd-order, so drive the LO at the modulator's recommended level and no more.

Next, the image. The IQ modulator's image is simply the unwanted sideband; its suppression depends on amplitude and phase balance. Write the amplitude imbalance as a linear ratio r and the phase quadrature error as Δφ (radians); the ideal image rejection (dB) is approximately 20log[r / √(1−2r·cosΔφ+r²)]. When r≈1 this can be roughly remembered as 20log(Δφ/2): a phase error of 2° (0.035 rad) gives about −35 dBc; 1°, about −41 dBc; 0.5°, about −47 dBc. Amplitude error must be judged separately: a 0.5 dB amplitude imbalance alone gives an image of only about −32 dBc; 0.2 dB, about −40 dBc. Both errors exist simultaneously — take the worse of the two, and keep 3–5 dB of margin for temperature drift. So when I write internal specifications I usually put "image better than −35 dBc across temperature," which corresponds to amplitude within 0.3 dB and phase within 1.5°. To push further you need on-chip calibration or digital predistortion — potentiometers cannot hold against temperature drift.

You must also watch LO leakage: it sits at the LO frequency, and if the LO happens to fall inside another segment's passband, no amount of front-end filtering will stop it — you can only push it down with IQ DC offset trimming (good for 10–15 dB) plus frequency-selective filtering.

One more warning: when the LO is 1.1 GHz and the IF is 1.2–1.4 GHz using the sum sideband, the difference product 0.1–0.3 GHz lands right inside the 100–500 segment's passband. This is the classic trap of "one LO feeding two segments at once" — you must separate them with segmented filtering plus channel enable staggering.

One final rule of experience: this spur table must be computed together with the filter rejection — spur level equals the modulator output spur, minus the filter rejection at that frequency, minus the downstream amplifier's out-of-band gain compression. As long as the result is below −50 dBc, acceptance will basically not go wrong.

2.5 Board splitting: three boards, each with its own duty

95 × 55 mm cannot hold it all, so split into three boards. The first is the RF main board, about 85 × 45 mm, four layers, carrying the IQ modulator, filter bank, attenuator, two amplifier stages, coupler, and detector. This board is pure RF: no digital power trace is allowed to cross the RF zone. The second is the digital and frequency-source board, about 80 × 40 mm, six layers, holding the FPGA, DAC, oscillator, PLL, VCO, and power. It is the noise monster; it is separated from the RF main board by standoffs and a shielded cavity, with only the IF, LO, and control lines crossing between them — all as RF lines or shielded ribbon. The third is the interface board, very small, carrying the RS-422 transceivers, DC jack, I/O-CTRL connectors, ESD parts, and reverse-polarity and surge protection on the power input.

Stack the three boards: the RF main board mounts against the cavity floor, the digital board sits above it, with a shielding partition between them. The benefit: the RF main board's ground plane stays intact, never chopped up by return currents from the digital board. Dual channel then means two RF main boards mounted back to back in two compartments of the same cavity, separated by an aluminum wall that must rise at least 3 mm above the board surface at the SMA centers — otherwise the two channels will couple. Measured isolation reaches 60 dB or better.

A newcomer asks: why not two boards, folding the interface onto the digital board? The answer is serviceability and reusability. The interface board is the most likely casualty of plugging, unplugging, and field wiring; as a separate small board, when it dies you replace a 20-yuan part. And when the next project changes the interface from RS-422 to CAN or gigabit Ethernet, only this one board changes — the RF and digital sections stay untouched. That is the platform mindset, and it should become your habit when building product lines.

Chapter 3: Core Component Selection — One Part at a Time, with Part Numbers and Pitfalls

3.1 The power amplifier: small power, but the selection logic is identical to a 100 W PA

First, a matter of mindset: don't grab just any Gain Block because you only need 5 dBm. The selection logic for a power amplifier is the same whether it is 5 dBm or as much as 100 W: first the in-band gain and gain slope, then P1dB and linearity after back-off, then PAE and thermals, then stability, and finally consistency and procureability. Let me walk through this logic on a low-power example — later, when you build tens or hundreds of watts, copy it verbatim.

For the final stage I chose a medium-power wideband GaAs or GaN driver-type PA with these requirements: 100–3000 MHz, small-signal gain 15–20 dB, P1dB not below 23 dBm, OIP3 not below 35 dBm, in-band gain ripple within ±1.5 dB (flattened by an external equalizer), single 5 V / 8 V supply, current under 150 mA (under 1 W at 5 V — fits the power budget). There are quite a few domestic options at this tier: Sichuan Yifeng, Beijing Huixin Tong, Borui Jixin, Yaguang Technology, and CETC 13th Institute all offer wideband PA die in the 0.1–3 GHz, sub-1-watt class. Huixin Tong's HXTA3036T89 — a 0.01–3 GHz wideband part with 20 dB gain, P1dB 18.5 dBm, OIP3 32 dBm, 5 V at 50 mA — is well suited for the front-end stage; for the final stage, move up to a 0.1–3 GHz, 1-watt-class GaN or GaAs part.

If the customer later raises the power spec from 5 dBm to 30-40 dBm — common in jammer applications — go look at Ruichuang Microwave's D110X series: the D1102 is 0.03–3.2 GHz, 10 W, 28 V, gain 9–13.7 dB, Psat 40.8 dBm, efficiency above 52.5%, internally matched at the input, plastic DFN package, usable directly; the D1103 for 15 W, and the D1105 (0.01–1.2 GHz, 10 W, 59% efficiency) for higher efficiency below 1.2 GHz. For the 100 W class, look at Xinbaité's CBNTL100 (20–3000 MHz, 28 V, 100 W, gain 19.5 dB, 70% efficiency) or Borui Jixin's internally matched wideband GaN. Keep these on file — when the customer raises the power, you won't have to relearn anything.

Huashi's path on chips is a bit different: they own three process platforms — GaAs, GaN, and InP — with self-developed GaN HEMT power density reaching 6 W/mm, covering 0.1 to 100 GHz, plus their F-RISE packaging platform and a 3,000 m² packaging and micro-assembly line of their own. What does that mean? If the shelf die you can buy has an in-band slope you don't like, or you need it packaged with your own pinout definition, you can go straight to Mr. Li (18688706870) to discuss custom or semi-custom: die, carrier, wire bonding, and housing assembled into a module for you, delivered together with full-temperature S-parameters and power sweep data. Shelf civilian chips solve "available"; this kind of collaboration solves "good" and "stable." On major programs — satellites, airborne platforms — only the latter holds up.

3.2 The LNA: why does a transmit source need one?

Someone will ask: this is a transmit source — what do we need an LNA for? Two uses. First, the detector branch: the coupled signal is only about −10 dBm, and before it goes into the detector and ADC for power closed-loop control, an LNA in front lifts the dynamic range and resolution so that power readings hold ±0.5 dB. Second, future transceiver variants: this module can easily be turned into a transceiver; then the receive path needs an LNA, and selecting it once now means no board respin later.

Domestic LNAs at 0.1–3 GHz are the most mature category in domestic RF: many models, low prices, good specs. A 0.1–3 GHz part with 21.5 dB gain, 1.0 dB noise figure, P1dB 15 dBm, 5 V at 23 mA in a 2 × 2 mm plastic package; for higher gain, a 0.1–3.5 GHz part with 29.5 dB gain, 0.9 dB noise figure, P1dB 19 dBm; for spaceborne or harsh environments, the ceramic-packaged YLN109-0004A2C — 0.1–3.5 GHz, 31.5 dB gain, 0.7 dB noise figure. Parts in the 0.01–3 GHz class with 1.0–1.2 dB noise figure, P1dB 18.5 dBm, and OIP3 32 dBm are also very usable. And there are wideband LNAs — that 1 MHz-to-4 GHz part with 2.0 dB noise figure at 2 GHz, OIP3 21.9 dBm, 3.3 V at 37 mA, on-chip bias — whose external circuitry is the simplest of all.

Again: this module is a transmit source and the main chain needs no LNA. I include it for two reasons. One, the coupled detector branch: with 16 dB of coupling after the final stage, the coupled port sits at −23 to −13 dBm. The log detector's input floor is around −50 dBm — apparently enough, but add the detector front-end limiter loss of 1 dB, trace loss of 0.5 dB, and a 6 dB accuracy degradation at the low-power end, and the loop starts jittering at low power settings. So put a 15–20 dB LNA after the coupled port to lift the whole detection range — then the loop is stable. Two, receive derivatives: the same board, slightly modified, becomes a receive exciter or reconnaissance front end — and then the LNA is the main chain's first stage, and its noise figure directly sets system sensitivity.

Selection looks at four things: noise figure, gain, in-band flatness, OIP3. For 0.1–3 GHz with 1.5–2 dB noise figure, 20 dB gain, and OIP3 above 20 dBm, Huashi's shelf has plenty. A 0.1–6 GHz part (noise figure ~1.5 dB, gain ~20 dB), a low-band part (noise figure under 1 dB, gain above 30 dB), and an 8 V, 120 mA part (P1dB 26 dBm, suited for a high-linearity front end) — all of these wideband LNA die work.

Three pitfalls. First: the LNA fears mismatch above all; the moment the input VSWR degrades, the noise figure turns ugly — so tune the input for minimum noise matching, not for 50 Ω; the two often differ by a few tenths of a dB of gain. Second: there must be limiting or switching protection in front of the LNA — your board's output end doesn't fear incoming signals, but the coupled port and test ports will get mis-connected. Third: the LNA's out-of-band gain will amplify big low-frequency signals and compress the following stage — the gain above 3 GHz must be held down; ideally put a highpass section between stages.

For LNA selection watch three numbers: noise figure, P1dB (or IIP3), and in-band gain roll-off. At 0.1–3 GHz, wideband LNA gain typically drops 2–3 dB at the high end (set jointly by device fT and the matching network), while at the low end gain bulges easily because the bias inductance self-resonates. That bulge is the precursor of self-oscillation; later in the matching chapter I will show how to flatten it with resistive feedback.

3.3 The filter bank: segmentation, switching, three bands

The frequency planning fixed three segments: 100–500, 500–1500, and 1500–3000 MHz. These three are built differently. LTCC filter devices are mature in this range: small, temperature-stable, insertion loss 1.5–2.5 dB, more than 45 dB rejection below 1 GHz out of band. But LTCC has one unavoidable trap: its out-of-band rejection relies on in-package parasitic notches, which drift by tens of MHz between lots — so demand measured S2P files from at least three production lots, and use the worst curve for the system spurious budget; never use the datasheet's typical curve. Also, LTCC is multilayer ceramic and its pads are the RF ports: follow the vendor's reflow profile exactly — a ramp above 3 °C/s will crack it — and never put vias in the pads or run heat-spreading copper under them. That is the most common volume-production assembly-line mistake.

For 100–500 MHz, lumped LC is the best fit: a 5th- to 7th-order elliptic or Chebyshev, with Sunlord wirewound or multilayer high-frequency inductors (Q above 40) and capacitors. The key in this band is the low end: at 100 MHz an inductor's impedance is only about j60 Ω and a 100 pF capacitor is j16 Ω — component values are large, parasitics matter less, and it's easy; the hard part is high-end rejection — holding 40 dB above 1 GHz takes two extra notches. For 500–1500 MHz, use distributed elements: a coupled-line or hairpin bandpass on Rogers 4350B (εr 3.48, loss tangent 0.0037), three to four sections, insertion loss held within 1.2 dB, with more than 35 dB rejection above 2 GHz.

What about switches? For 0.1–3 GHz with insertion loss under 1 dB, isolation above 30 dB, and P1dB above 25 dBm, domestic GaAs SP3T parts or combinations of SPDTs all work — GaAs switch chips are cheap in volume. For more cost-driven civilian projects, SiGe or SOI switches are usable too, but watch their voltage handling and P1dB: SOI switch P1dB is generally only in the low 20s. Placed at a low-power point — after the attenuator, before the final stage — fine; placed after the final stage, it will burn.

3.4 The core FPGA: control, DDS, predistortion — one chip does it all

The FPGA on this board must do: a 48-bit DDS phase accumulator; IQ data shaping for the 200 MHz bandwidth (noise or sweep waveforms); parallel control of the 31.5 dB attenuator; DAC control of the VGA; reading the detector ADC and closing the power loop; SPI configuration of the PLL; the RS-422 UART and cascade repeating; and temperature-compensation table lookups. The logic volume is not large, but you need enough multipliers — a 200 MHz FIR for predistortion or pulse shaping takes a few dozen DSP blocks.

Three tiers of selection, by project grade. For civilian and commercial aerospace: about 50k LUTs, industrial −40 to 100 °C, free PDS toolchain, BGA256 package, friendly price; step up to the PGL100H if predistortion logic overflows. For demanding spaceborne applications: the 326k-LUT part with 840 DSPs and 16.4 Mb of BRAM, with the highest acceptance in military-aerospace programs; Huashi's spaceborne products use exactly these high-reliability FPGAs, and radiation-tolerant variants can be selected — for configuration details go to Mr. Li (18688706870), who can provide irradiation and temperature-lot data. For projects squeezed to the last cent: a small-logic device is enough for everything except the main datapath (I2C to control a DAC, read an NTC); many counter-drone projects get by with one small FPGA plus one DAC for temperature compensation.

Three pitfalls of FPGA selection, stated in advance. One: the development tools — the syntax and IP cores differ from Vivado; the PLL, FIFO, and DDR IP all have to be relearned; reserve two weeks. Two: domestic FPGAs' PLL jitter is 20–50 ps worse than the equivalent Xilinx part — so the DAC's clock must go from the external oscillator through the clock chip straight to the DAC, never out through the FPGA's internal PLL. Three: the configuration chip — for a domestic FPGA's SPI Flash, the boot-mode pins must be pulled up or down per the manual; never leave them floating.

3.5 DAC and clock: the source of all spurs

Choose a 14-bit, 1 GSPS-class DAC; or a DDS-integrated DAC (on-chip 1 GSPS 14-bit DAC plus DDS, serial IQ input, narrowband SFDR above 80 dB), which eliminates the NCO in the FPGA and cuts both logic volume and power. If money is short, go 12-bit — but a 12-bit SFDR is about 60–65 dB: enough for a jammer, not enough for an instrument-grade source.

The clock is the most expensive part in this module, and the one you may not economize on. The DAC's sample-clock phase noise directly sets the output's phase noise and spurious floor. Use a 100 MHz TCXO (±0.5 ppm, −40 to 85 °C) as reference, into a domestic clock chip or PLL (Chengdu Huawei's clock drivers, or the 13th Institute's clock devices) to multiply up to 1 GHz with jitter in the 100 fs class. The oscillator's supply must have its own dedicated LDO; its routing stays at least 5 mm from the FPGA and any DC-DC, over a complete ground plane, with absolutely no signal line passing under the crystal — I will yell about this again in the layout chapter.

3.6 PLL, VCO, and LO: the trade between hop speed and phase noise

The LO must hop between 1.1 and 4.4 GHz and lock within 10 µs. The approach: a wideband fractional-N PLL plus a wideband VCO, or a single integrated-VCO wideband PLL. A part with 8 GHz max input and a normalized noise floor of −227 dBc/Hz, paired with a suitable VCO, is a proven combination; integrated solutions also exist (from the former 24th Institute lineage and other frequency-source chips). Three cautions. One: fractional-N spurs — the ΣΔ modulator's spurs can land in-band, so either pick a part with spur cancellation or suppress them in the loop filter. Two: loop bandwidth — we want 10 µs lock, which needs a loop bandwidth above 100 kHz, but opened wide, the VCO's phase noise escapes; compromise at 50–80 kHz for a lock time around 15 µs, which is enough. Three: align the VCO's tuning voltage range with the PLL's charge-pump output (usually 0.5–4.5 V); if they don't align, add an op-amp level shifter — don't force it.

3.7 Detection, ADC, and supply monitoring

For power detection use a domestic log or RMS detector in the 0.1–3 GHz class with 40 dB dynamic range and ±1 dB accuracy — Huashi's shelf covers it. Behind it, a 12-bit ADC — a domestic SAR ADC at 1 MSPS or better — feeds the FPGA for the power loop. Temperature comes from an NTC or a digital temperature sensor; current from a sense resistor plus a zero-drift op-amp. The I_ALARM and T_ALARM outputs are 3.3 V logic from FPGA GPIOs through a driver; remember the 22 Ω series resistors and TVS diodes — once the interface cable gets long, ESD and surge come in right here.

3.8 Passives and connectors: the cheapest parts, the easiest to cause trouble

Capacitors: C0G/NP0 only. Never use X7R or Y5V in matching arms — X7R can lose more than 30% of its capacitance with voltage, temperature, and frequency; your match looks fine at room temperature and detunes when hot. A 0402 100 pF C0G self-resonates around 1.5–2 GHz; at 3 GHz it is already an inductor. So near 3 GHz, bypass with 10 pF and 100 pF in parallel to stagger the self-resonances. Inductors: Sunlord multilayer high-frequency or wirewound types — but read the Q at the datasheet's test frequency: many vendors specify Q measured at 100 MHz; used at 2 GHz, the real Q may be only a third of it.

Board-to-board and external connections: RF on SMA — pick the "removable stainless steel, 500 mating cycles" grade, not the few-mao brass gold-plated kind, whose VSWR degrades after three matings. The DC jack should be a threaded locking aviation connector or a 2.54 mm dual-row header with latch; I/O-CTRL on 2.0 / 2.54 mm socket strips. RS-422 transceivers: an isolated, 20 Mbps, wide-body SOIC16, −40 to 125 °C part or its equivalent — long cascade lines must be isolated; ground potential differences will kill non-isolated transceivers.

Chapter 4: Wideband Matching — the Real Craft of This Board Lives in This Chapter

Everything so far has been selection; the judgment behind selection rests on the craft in this chapter. Read this chapter three times — it is worth more than ten papers.

4.1 First, build a concept: wideband matching is not "pushing S11 down"

Narrowband matching: with a Smith chart you add a series inductor, a shunt capacitor, drag one frequency's impedance to 50 Ω, done. Wideband doesn't work that way. The essence of wideband matching is a multi-section network that keeps the reflection coefficient within spec at every frequency along the long trajectory from 100 MHz to 3 GHz, while the network's gain slope also stays flat. These two requirements are usually at war: the better the match, the more "frequency-picky" the network and the steeper the gain slope; flatten the slope and the edge-band return loss sags. So the first mindset shift of wideband matching is: you are not matching a point — you are shaping a curve.

The second mindset shift: accept "non-ideal." Achieving S11 better than −10 dB across the whole 0.1–3 GHz band is neither easy nor necessary in practice. Our internal standard: in-band S11 better than −8 dB (VSWR under 2.2), S22 better than −10 dB (the output end matters more, because the filter follows and a mismatch there ruins the filter's in-band ripple), gain flatness ±1 dB. Setting the spec where it can actually be built and still pass customer acceptance — that is the chief designer's job.

4.2 Four topologies and where each belongs

First: resistive feedback. The oldest and most effective trick in wideband amplifiers: an RC series feedback network from drain (collector) to gate (base). It does three things — pulls input and output impedance toward 50 Ω, flattens the gain (as high-end gain falls, the effective feedback drops, giving automatic compensation), and dramatically improves stability. The cost: noise figure worsens by 0.5–1.5 dB and output power drops 1–2 dB. We rarely use it on the final stage (we begrudge the power), but always in driver stages and VGAs. How to size it? Back-derive from the gain target: for 15 dB of gain, the feedback resistor typically runs 200–600 Ω and the capacitor 0.5–2 pF (setting the frequency where feedback starts to act, fc = 1/(2πRC)), then simulate and sweep. Rule of thumb: for a 0.1–3 GHz GaAs device, Rf = 300 Ω and Cf = 1 pF is a very good starting point.

Second: source (emitter) degeneration. Put a small inductor (or an inductor paralleled with a resistor) in the source leg. At the low end the inductor's impedance is small, feedback weak, gain high; at the high end impedance large, feedback strong, gain low — exactly the inverse of the transistor's native "high-end gain droop." At 0.1–3 GHz, a source inductance of 0.5–2 nH (a few millimeters of microstrip is enough) lifts high-end gain by 2–4 dB. Caution: source inductance degrades stability, especially at the low end — you often need to parallel it with a few-ohm to ten-ohm resistor as damping, trading a bit of gain for unconditional stability.

Third: multi-section RC/LC equalization (lossy equalization). This is the main weapon against gain slope. The principle is simple: a branch of series resistor with shunt inductor — at low frequency the inductor shorts the signal into the resistor (attenuation); at high frequency the inductor opens and the signal passes — a positive-slope equalizer that exactly compensates an amplifier that is "high at the low end, low at the low end." Reverse it — series capacitor with shunt resistor — for a negative slope. Over 0.1–3 GHz I use two sections: one to compensate 100–500 MHz (R = 15 Ω, L = 33 nH, 3–4 dB of attenuation at the low end), one for 1.5–3 GHz (R = 10 Ω, L = 6.8 nH). Lossy equalization costs 3–5 dB of gain, so the link budget must reserve it in advance — which is exactly why the final stage was sized to 23 dBm in the power chapter.

Fourth: the distributed (traveling-wave) structure. Treat the input and output capacitances of several transistors as the capacitances of an artificial transmission line, joined by short line segments; bandwidth can reach tens of times. But the die area is large and efficiency low, so it is normally done inside an MMIC. The wideband chips you select are internally distributed or feedback types that come pre-matched; what you design is "the matching around the chip," not "the matching from the transistor up." Recognizing this point saves you many detours.

—— Huashi (Shenzhen Huashi Digital Technology Co., Ltd.), tel: 186 8870 6870

客服中心
联系方式
18688706870
0755-29210759
- 销售
微信18688706870
seo seo