扫码咨询

新闻动态

首页 >> 新闻动态 >> From 11.5 GHz Sweep to 92-96 GHz Output: Multiplier Chain, PA Back-off and I/Q Balance, the Three Bottlenecks of W-band Volume Production
详细内容

From 11.5 GHz Sweep to 92-96 GHz Output: Multiplier Chain, PA Back-off and I/Q Balance, the Three Bottlenecks of W-band Volume Production

An external IF sweep at 11.5 to 12 GHz becomes a 92 to 96 GHz RF output after an eight-times multiplier chain. The receiver is a zero-IF I/Q architecture, and its DC to 1 GHz band is the output bandwidth, not the RF bandwidth. On a block diagram every line item looks as if it already has a matching part number. In volume production, three items still decide yield: the spurious content and in-band flatness of the multiplier chain, the back-off level and thermal behaviour of the final power amplifier, and the amplitude and phase balance of the two I/Q paths. Shenzhen Huashi Digital Technology Co., Ltd. designs, manufactures and sells its W-band front-end assemblies together with the LNA, PA, multiplier, mixer and detector die behind them. This article books the whole chain in that order: first the allocation of system numbers to individual stages, then part selection on one and the same fixture, and finally the closing of the loop through schematic nodes, layout dimensions, assembly cards and production limits.

1. Split the spec into numbers each stage can own

The point that misleads newcomers most is worth stating first. The external IF input is an 11.5 to 12 GHz sweep, and the eighth harmonic of it becomes the 92 to 96 GHz RF signal. The DC to 1 GHz figure belongs to the I/Q baseband output. Transmit power, receive noise figure, conversion gain and I/Q balance are not four independent line items either: on one chain they compete for the same margins. Allocate those four numbers to individual stages first, and the schematic, the layout, the housing, the fixture and the acceptance plan all have somewhere to sit.

Receive chain: where the 4 dB goes. Take a waveguide transition at 0.25 dB, a receive pre-filter and the necessary protection at 0.30 dB in total, a first-stage LNA (Huashi HS-LNA-75110C, 75 to 110 GHz, 23 dB typical gain, 2.8 dB noise figure at 90 GHz, output P1dB around +1 dBm, 1 V drain bias and about 33 mA total drain current), a 0.20 dB inter-stage transition, a second stage with 10 dB gain, a passive mixer with 12 dB conversion loss, then IF amplification. Run Friis with linear noise factors. The transition and the protection add up to 0.55 dB, a linear loss of about 1.135; the LNA at 2.8 dB is 1.905, and the first two blocks combine to about 2.163. The 23 dB first-stage gain is 199.5, so the 0.20 dB inter-stage loss contributes about 0.115/199.5, a 4 dB second stage contributes about 1.585/199.5, and the 12 dB mixer loss with a following 3 dB block contributes, after division by 199.5 times 10, roughly 0.008. The total noise factor comes to about 2.177, which is 3.38 dB.

That leaves 0.6 dB of margin against 4 dB, before connector repeatability and temperature are counted. The internal control limit for volume production is therefore 3.6 dB, not 4.0 dB. Anyone who treats 3.9 dB as "passing with margin" will pay for it in production. If the front-end transition and protection together cost 0.3 dB more, the system noise figure moves by about 0.3 dB; if the first-stage gain is 3 dB lower, the later stages contribute a little more, and the result sits right on 4 dB. This is why the waveguide probe, the input ribbon bond, the protection device and the LNA bias matter far more than swapping in a mixer that looks stronger on paper. Every part ahead of the LNA is part of the noise factor, not just a pipe the signal passes through.

Transmit chain: +18 dBm is not one amplifier's job. The numbers 92 to 96 GHz, +17 dBm minimum and +18 dBm typical mean that no unit in the batch, anywhere in the band, at any rated temperature, may fall below the threshold. Assume 1.2 dB of loss between the final stage and the waveguide flange: the die must deliver at least +19.2 dBm into its optimum load. Add assembly spread, temperature drift and 0.3 dB of test uncertainty, and the on-board operating point should be designed at +20 to +21 dBm. The part choice then becomes clear. Huashi HS-PA-9296G25 is a 92 to 96 GHz GaN PA rated at 15 V with 180 mA quiescent current, 18 dB small-signal gain, 23 dBm saturated output and 8 percent power-added efficiency in continuous wave, a die of 3.68 by 1.34 by 0.05 mm and a typical gate voltage near -1.5 V. Huashi HS-PA-9296G32 covers the same band at 15 V with a typical gate voltage of -1.5 V, 14.5 dB small-signal gain, 30 dBm output and 15.5 percent power-added efficiency, in a die of 3.4 by 1.43 by 0.05 mm. The first is the conservative choice for a +18 dBm output in the first build. The second has far more headroom and suits a higher-power derivative, but 30 dBm cannot be treated as a result that arrives once the module is boxed.

Run the thermal sum for HS-PA-9296G25. At +21 dBm on board, about +3 dBm of input, 15 V drain and 200 mA of operating current, DC power is 3 W and net RF output is about 126 mW, which by definition gives a power-added efficiency near 4 percent. That does not contradict the 8 percent in the data sheet at saturation: a back-off operating point is not usually the efficiency peak. The die itself dissipates close to 2.87 W, and with the driver and multiplier chain added, housing cooling is no longer a matter of bolting on any aluminium shell. With a 30 dBm class device, the same 15 V and 200 mA assumption gives a higher dissipation; ask the vendor for the current at the intended back-off point instead of estimating the supply and case temperature from the Psat line.

Coherence and bandwidth: same source does not mean no phase problem. Four gigahertz against 94 GHz is about 4.3 percent, which makes the waveguide transition, the LNA match, the multiplier output, the PA output, the LO drive and the I/Q hybrid all wideband problems. Design reviews should cover 92, 93, 94, 95 and 96 GHz at minimum, and ideally extrapolate to 91.5 and 96.5 GHz to see the edge slope. Lighting up at 94 GHz alone is not a 4 GHz bandwidth. Frequency multiplication of the 12 GHz source degrades phase noise by 20log10(8), about 18.06 dB, in the ideal case. If the 12 GHz source sits at -100 dBc/Hz at 100 kHz offset, expect roughly -82 dBc/Hz after multiplication, before the multiplier chain amplifiers, the supplies and the reference spurs add their own noise. What the front end can control is the source interface match, the multiplier input band-pass, the low-noise bias, the LO buffer isolation and the discipline that keeps digital clocks from modulating the LO. A low-noise regulator cannot repair a noisy source, and a poor range sidelobe should not be blamed on the LNA.

2. Architecture: multiplier, conversion and I/Q trade-offs

The multiplier chain. This is where "there is output" is most easily mistaken for "performance is acceptable". Huashi HS-MUL8-8096H is an 80 to 96 GHz active x8 multiplier with 3 to 7 dBm input, 12 dBm output and 15 dBc harmonic rejection. HS-MUL8-8096L covers the same band and input range with 3 dBm output and 15 dBc rejection. HS-MUL8-7090L covers 70 to 90 GHz with 3 to 7 dBm input and 2 dBm output, HS-MUL8-7090H covers the same band with 11 dBm output, and HS-MUL9-8010A is an 80 to 100 GHz x9 multiplier with 14 to 17 dBm input, 2 dBm output and 30 dBc rejection. Mind the definition first: a catalogue harmonic number is not automatically the X7 and X9 spurious figure you need, and it says nothing about what survives the die pads, the ribbon bonds, the DC blocks, the filter and the waveguide transition. A 12 dBm output also does not mean 12 dBm arrives at the PA input; every transition and filter takes a few tenths of a decibel. The first build is better served by a driven multiplier such as HS-MUL8-8096H, not because it is the most powerful, but because it leaves drive margin for the stage behind it. Assess it by sweeping 11.5, 11.75 and 12 GHz while also sweeping input power from 0 to 8 dBm, and record the output at 92, 94 and 96 GHz together with X7, X9 and LO leakage. The acceptance line should not read "output at least some value" but rather: within the specified input power window, multiplier output flatness no worse than 1.5 dB, X7 and X9 after filtering no higher than -35 dBc at the transmit port, and input return loss that does not detune the preceding stage. If only a 3 dBm output sample is at hand, do not reject it outright: adding a 20 dB driver behind it may buy better in-band shaping and spurious control.

Mixers and I/Q. Huashi HS-MIX-70110A covers 70 to 110 GHz with IF up to DC to 35 GHz, 14 dB conversion loss, 8 dBm LO drive and 2 dB LO-to-RF isolation. HS-MIX-70110B covers the same band with 12 dB conversion loss, 13 dBm LO drive and 25 dB isolation. A word of warning for newcomers: a single isolation figure in a catalogue refers to one port pair only, and it is not image rejection. The B version saves 2 dB of conversion loss but asks for 5 dBm more LO drive, and at W-band those 5 dBm usually mean recalculating the LO buffer, the dissipation, the isolation and the cavity crosstalk. A zero-IF receiver is also more sensitive to LO leakage back into the antenna and to mixer DC drift, so the choice is not made on 2 dB of loss. Mount both devices on the same evaluation board and measure the LO power plateau, the I/Q amplitude and phase tracking against frequency, the IF DC drift and the temperature behaviour. With two single mixers and a 90 degree hybrid, measure the hybrid separately before anything else.

The same band supports zero-IF, low-IF and image-reject architectures, and a block diagram should not be copied blindly. Huashi HS-IQM-9296 is a 92 to 96 GHz image-reject I/Q mixer and serves as a structural reference. HS-UPC-9296 integrates a mixer and a medium power amplifier to convert 2 to 6 GHz IF up to 92 to 96 GHz, which makes it a useful comparison for the transmit path. HS-DNC-9396 converts 93 to 96 GHz down to 3 to 6 GHz and is a natural superheterodyne option. Zero-IF saves filtering and suits a common-source coherent design, but DC offset, 1/f noise and LO leakage hurt most. Low-IF moves the signal away from DC and makes the analogue chain easier, at the price of ADC speed, image filtering and LO planning. When the project genuinely needs a DC to 1 GHz baseband, zero-IF is the reasonable route, provided the hardware carries DC offset compensation and I/Q calibration.

Filter allocation. Filter specs must reach the level of a measured curve. On the transmit side, start with a compact waveguide band-pass of 2 to 3 dB loss after the multiplier and ask for no more than 2.5 dB in band, 0.8 dB of ripple, and 30 dB of rejection below 91 GHz and above 97 GHz as a first discussion point, with the final rejection set by the spurious budget. The receive pre-filter must not copy that value: 1 dB of loss there costs the system noise figure about 1 dB. Design the pre-filter for 0.3 to 0.5 dB of loss with a wider transition, and put the narrow rejection behind the LNA. Waveguide filters offer high Q and power handling at the cost of size and money; substrate integrated waveguide and suspended stripline integrate easily with thin-film circuits but are more sensitive to edge roughness and lid and dielectric tolerances. The decisive question is not the process name but whether the vendor can supply measured S-parameters, group delay and the spread of at least five samples.

3. Control, data conversion and device substitution

The controller does not need to be large. The digital work here is small: SPI configuration, DAC codes, ADC sampling, temperature read-out, power-up sequencing, fault latching and calibration storage. Huashi HS-FPGA-690T offers roughly 690k logic cells, 3600 DSP slices and 20 GTH transceivers, which is far more than a single-channel front end needs. A mid-size device from the Huashi HS-FPGA-N or HS-FPGA-E family, or an HS-MCU-32 series part with a small CPLD, fits better. What must actually be confirmed is the operating temperature range, the configuration memory, the I/O levels, the start-up current, the tool chain version and the long-term supply, not the logic count. A large FPGA converts switching noise, extra power layers, cooling and firmware maintenance into cost at volume.

Choose the ADC and DAC by real bandwidth. Each I/Q path carries 1 GHz of analogue bandwidth. Keeping the baseband to 1 GHz means at least 2 GSPS by Nyquist, and 2.5 to 3 GSPS is the practical starting point. Huashi HS-ADC-14B2000, HS-ADC-14B2600 and HS-ADC-14B3000 are 14-bit parts at 2.0, 2.6 and 3.0 GSPS with JESD204B interfaces. Start interface evaluation at 14 bits and 2.5 to 3 GSPS, then check the analogue input bandwidth, the signal-to-noise ratio, the spurious-free dynamic range, the power and the JESD link. If the system really needs only a few hundred megahertz of processing bandwidth, do not move to 3 GSPS to "cover 1 GHz": place the analogue anti-alias corner at 0.9 to 1.0 GHz and both the sample rate and the digital load fall. A DAC that only generates a low-frequency modulation ahead of a 12 GHz LO should be judged on its analogue bandwidth, its images and its clock phase noise, not on its bit count.

Substituting the core PA. Four things are most often missed. First, the input drive differs: a larger die has more input capacitance, and a driver sized for the old part may compress before the new one does. Second, the gate negative supply range and the power-up sequence differ: get the gate sequence wrong and the drain current runs away the moment the drain supply arrives. Third, the optimum load differs: the existing match was built around the old die's load-pull result, and a new S22 cannot be fixed by trimming bias. Fourth, the thermal design differs: identical package outline does not mean identical backside metallisation, die thickness or thermal resistance. The habit that pays is to give each candidate its own pads and its own matching area on the board, even at the cost of a few square millimetres. Saving that area and forcing two supply chains onto one copper island leaves only one choice later, and then there is no second source at all.

Huashi HS-PA-8810A is specified for 88 to 100 GHz with 16 dB gain, 22 dBm saturated output and 21 dBm P1dB at 4 V and 335 mA in a 3.40 by 2.10 by 0.07 mm die. It suits a low-voltage, low-dissipation driver or a lower-power final stage, and it is not a direct replacement for a 15 V GaN output stage. Huashi HS-PA-9296G28 covers 92 to 96 GHz with 20 dB gain and 26.5 dBm saturated output and is the GaN candidate closest to this band. HS-PA-9098G31 spans 90 to 98 GHz with 20 dB gain and 29 dBm saturated output, with more headroom but with matching, cooling and supply all to be redone. Before keeping or discarding an existing match, compare gate voltage, drain voltage, quiescent current, S-parameters, load-pull data, die thickness, backside metallisation, thermal resistance and pad drawing. Comparing only band, gain and Psat is the most common and the most expensive substitution error.

4. Schematic discipline: five numbers at every stage

Every stage on a schematic review must carry five numbers beside it: power, noise, gain, compression point and supply. On the transmit side, start at the SMA: 12 GHz input from 0 to +8 dBm, with the multiplier input nominally at +3 dBm; multiplier on-board output estimated at +9 to +12 dBm; +7 to +10 dBm after the band-pass; 18 to 22 dB of driver gain with the output held at +10 to +13 dBm; the final stage checked against that range; +20 to +21 dBm on board at the final stage and +18 to +19.8 dBm at the waveguide flange. On the receive side, start at WR-10: with a -40 dBm test input, the 0.55 dB front-end loss puts about -40.55 dBm at the LNA, 23 dB of gain brings it to about -17.55 dBm, the second stage to about -7.55 dBm, and the 12 dB mixer loss leaves about -19.55 dBm, so the IF section needs 20 to 30 dB before a 1 GHz bandwidth acquisition stage. These are continuous-wave reference points, not limits; written this way the chain shows at a glance that the LNA never compresses and how much IF gain the mixer needs.

Draw the bias network in three parts: RF decoupling at the die pad, low-frequency energy storage near the die, and filtering and current limiting at the cavity feed-through. At W-band both gate and drain need 100 pF and 1000 pF class capacitors close to the die, but the self-resonance of each package must be confirmed; a part marked 1000 pF is not necessarily a capacitor at 94 GHz. Huashi HS-PA-9296G25 also requires external 100 pF and 1000 pF capacitors on both gate and drain bias, with the bottom ground serving as RF and DC ground at once. The reason is not superstition about extra capacitors: it is to stop millimetre-wave energy from travelling down the supply lines into the bias network and being amplified elsewhere. Keep the PA drain trace short and wide, keep the return current away from the LNA ground, and bring the gate supply up softly before the drain, with the reverse order at shutdown. No control software may bypass the hardware undervoltage and overcurrent protection.

5. Wideband matching: impedance target first, topology second

State the impedance target before discussing topology. A PA output usually uses a series high-impedance line acting as an inductor with a low-impedance shunt section to form a low-pass network, which also suppresses the second harmonic. A PA input can use two stepped-impedance sections with one tuning pad. An LNA input is matched for minimum noise, not for the prettiest S11. A mixer LO port is matched for the drive plateau, and its RF port for a compromise between conversion loss and return loss. On 100 ohm alumina microstrip, with an effective dielectric constant near 6, the guided wavelength at 94 GHz is about 1.30 mm and a quarter wavelength about 0.33 mm. On a 0.127 mm Rogers 3003 class soft board, with an effective dielectric constant near 2.3, the quarter wavelength is about 0.66 mm. Both figures are layout starting points only; the ribbon bonds, pads, ground vias, lid, adhesive layer and waveguide probe must all enter a three-dimensional electromagnetic simulation before the layout is frozen.

Ribbon bonds cannot be hand-waved either. A 25 micron wire, 300 microns long, is not a short circuit at 94 GHz. Two in parallel lower the inductance, but spacing, arc height and ground-return bonds all change the result. The process card should state RF bond length no more than 300 microns and arc height no more than 100 microns, with ground bonds paired on both sides of the signal wherever possible; if that cannot be met, change the layout rather than asking an operator to keep it short. Prefer two parallel ribbons at the PA output pad, a single short ribbon at the input with trim allowance, and restrict tuning sites to three states: empty, a 50 ohm thin-film resistor, or a capacitor or ribbon of defined size. Never let a solder blob of undefined length act as a component.

A lumped equivalent helps as a schematic starting point, provided it is not mistaken for a die model. Suppose the PA presents 15 - j10 ohm at 94 GHz near the target power and must be transformed to 50 ohm. A first guess of 0.16 nH gives +j94.5 ohm at 94 GHz, and 0.13 pF gives a susceptance near +j0.0768 S. Series inductor first, then shunt capacitor: normalised, 0.30 - j0.20 becomes 0.30 + j1.69, whose admittance is about 0.103 - j0.583; adding j1.54 leaves 0.103 - j0.957, still not unity. The hand calculation is the point: a single L and a single C will not complete a wideband match by guesswork, and such a branch tends to resonate too narrowly. Two distributed sections or a second real-to-real transformation are needed. In practice, split the 0.16 nH into two series contributions of about 0.08 nH each, and split the 0.13 pF into an adjustable short stub of about 0.07 pF plus about 0.06 pF of parasitic and phase compensation. Optimise with ideal lines to find the feasible region, then map the result onto line widths, line lengths and ribbons.

At +18 dBm the peak sinusoidal voltage on a 50 ohm line is about 2.82 V and the peak current about 56 mA; transform the die end down to 15 ohm and the current density near the same power is higher. The concern on an output line is conductor loss, current crowding, hot spots and ground return, not voltage standoff. Leave a replaceable region of 0.5 to 0.8 mm before the DC block, able to hold a 0.2 pF capacitor site, two shunt ground pads and a series jumper, and changeable to 0.1 pF, 0.3 pF or a plain ribbon. Never fit three things at once; change one variable per iteration. Set the output match target in three tiers: first S11 no worse than -8 dB, enough for stability and output power; second S11 no worse than -12 dB with 0.8 dB or less of loss across 92 to 96 GHz, which is acceptable for a prototype; third S11 no worse than -15 dB, loss no more than 0.5 dB and group-delay ripple within 15 ps. The last two tiers only mean something after the fixture is de-embedded.

The first rule at the receive input is not to force S11 to -20 dB but to make the noise match and the gain flatness hold. A typical path runs from WR-10 to microstrip, through a 0.3 to 0.5 pF DC block, a 50 ohm lead-in, a high-impedance short stub and a slightly lower impedance section into the LNA pad. The first build carries one shunt tuning stub and one series phase section, never two cascaded narrowband resonators ahead of the LNA. Take a 0.3 pF DC block: its ideal reactance at 94 GHz is about -j5.65 ohm, so it is not a transparent wire and it produces real phase shift and reflection in a 50 ohm system. Compare 0.2, 0.3 and 0.5 pF at the same position and the reactances are about -j8.48, -j5.65 and -j3.39 ohm. Choosing the value needs more than a self-resonance above 94 GHz: check the equivalent series inductance, the Q and the mounting ribbons at 94 GHz, because a packaged capacitor with two ribbons is often decided by the ribbons rather than the capacitor. Quantify protection as well. A limiter or diode with 0.10 pF of junction capacitance is about -j16.96 ohm at 94 GHz; even 100 microns away from the main line it can change the source impedance completely. When protection is unavoidable, prefer a device with an existing W-band reference design and simulate protection, DC block and match as one network.

A mixer is not a single-port amplifier: the RF, LO and IF ports interact through their reference planes, bias and terminations. Shape the LO path with a 50 ohm main line and a shunt open stub; give the RF port one short series line and one short stub to absorb the residual reactance; use a low-pass network at IF to block 94 GHz leakage. A useful first guess for the IF low-pass is 15 nH series with 3.3 pF shunt, which blocks RF strongly in theory, but a real 15 nH part is long past self-resonance at W-band, so it is only valid as a DC to 1 GHz passband starting value. Implement the actual RF choke with two high-impedance thin lines and two low-impedance stubs, and confirm that the first self-resonance sits well above the operating band. Start LO drive from the mixer recommendation: if the device asks for +13 dBm, do not run +7 dBm on board and then complain about conversion loss; equally, 3 dB above the recommendation is not automatically better and may raise LO leakage and I/Q imbalance. For in-line monitoring, sample with a 20 to 30 dB directional coupler and check that the detector works at the resulting level.

The multiplier input faces 11.5 to 12 GHz and its output faces 92 to 96 GHz, and the intermediate doubler and quadrupler nodes are not places to leave arbitrary line lengths. With a 12 GHz fundamental the seventh harmonic is 84 GHz, the eighth 96 GHz and the ninth 108 GHz. Since 84 and 96 GHz are only 12 GHz apart, an ordinary wideband low-pass cannot pass 96 GHz losslessly while strongly rejecting 84 GHz. The realistic route is balanced internal structure in the device plus an output band-pass or the waveguide high-pass behaviour, with the final stage and the load response doing the rest. Reserve a harmonic termination site at the multiplier output: a 35 to 60 ohm, 200 to 350 micron transforming line after the DC block, with an optional open or short pad beside the main line, left empty in the first build and connected as a compensation stub in the second only after measured harmonic phase data exist.

WR-10 waveguide is 2.54 mm wide and 1.27 mm high, with a TE10 cutoff near 59 GHz and single-mode operation from 75 to 110 GHz. An E-plane probe transition depends on the probe penetration, the back-short distance, the impedance taper and the ground-signal-ground return path together. Newcomers draw the probe and forget the lid, the cavity depth, the screws and the ground contact, yet a few tens of microns in any of them moves the whole return-loss curve. A simulation starting point, not a manufacturing release: 127 micron quartz or alumina carrier with 3 to 5 microns of metal; probe entering at the centre of the broad wall with 0.15 to 0.30 mm initial penetration; back-short distance swept from 0.25 to 0.55 mm; taper from about 25 to 35 ohm up to 50 ohm over 0.6 to 1.2 mm; ground vias or cavity walls within 100 to 200 microns of the signal edge. A first-round target of -10 dB return loss and 0.6 dB loss across 92 to 96 GHz is reasonable, with -15 dB and 0.35 dB as the stretch goal. When return loss is good but loss is high, look at surface roughness, probe contact, via return and cavity seams before touching line widths again.

Think in corner cases from the first layout: nominal, minimum line width with maximum dielectric thickness, maximum line width with minimum thickness, ribbon inductance low and high. Record worst-case S11, minimum and maximum S21, loss peak-to-peak and group-delay peak-to-peak. A nominal -20 dB S11 that degrades to -7 dB at the corners is not a production design, while a nominal -13 dB that stays below -9 dB at every corner with stable loss looks manufacturable. Tuning sites must map onto those tolerances: one series compensation site, one shunt stub site and one replaceable DC block at the PA output; exactly one main trim site at the LNA input so the noise optimum is not disturbed; one power-monitor coupling site at the LO port; common-mode and differential termination options at IF. Number every site on the silkscreen and define four BOM states: unpopulated, 0 ohm, capacitor and ribbon jumper, with the production file defaulting to unpopulated.

Simulation and measurement must close on the same reference plane. Calibrate the vector network analyser at waveguide ports before any W-band measurement, with the calibration plane ahead of the transition under test, or use a suitable substrate standard to move the reference plane to the die or board pad. After calibration, measure a known through and a load to confirm fixture repeatability. If three remounts of the same board scatter S21 by 0.2 dB, a 0.1 dB simulated difference cannot justify a layout change. Keep four data sets for every match measurement: raw calibrated data, gated and de-embedded data, both lid-clamped states, and cold and hot states. The lid state matters in particular, because a lid resting in place and a lid torqued to specification can be two different circuits; the mechanical drawing must state screw count, order and torque. Compare curve shape before absolute difference: a 1 to 2 GHz shift across the whole band points to dielectric constant, line length or cavity model; a high-end-only deviation points to ribbons, vias and package parasitics; a spike points to cavity resonance, ground discontinuity or a loose fixture; an extra 0.3 to 0.8 dB of loss points to surface roughness, transition contact and material loss.

6. Demo boards: turning vendor data into your own evidence

Never stop at one S21 curve. A PA board needs at least four data sets: small-signal S-parameters; output power, gain, current and power-added efficiency for input power from -10 to +15 dBm; power flatness at 92, 93, 94, 95 and 96 GHz; and thirty minutes of continuous power and case-temperature drift. An LNA board needs noise figure, gain, P1dB, IIP3 and drift from -10 to +50 degrees Celsius. A multiplier board needs input power sweeps, frequency sweeps, X7 and X9 content and LO leakage. A mixer board needs LO power sweeps, conversion loss, isolation, IF DC offset and I/Q amplitude and phase. Mark the reference plane on every curve. Without that layer of data the component values on the schematic are only vendor suggestions, and when a batch drifts there is nobody to blame.

Detectors deserve the same care. Huashi HS-DET-110A covers DC to 110 GHz with about 600 V/W sensitivity, and HS-DET-130B covers 70 to 130 GHz with about 3000 V/W, both over a -50 to -20 dBm dynamic range. With a +18 dBm transmit output, a 20 dB coupler delivers -2 dBm and is already outside the rated range of both; move to a 30 dB coupler, which gives -12 dBm at the coupled port, and the detector sits where it belongs. Coupling factor is therefore not an arbitrary choice: work out the detector operating point first, then the ADC range and the temperature drift.

7. Assembly, bring-up and production control

The first prototype uses a compartmented aluminium housing with a removable lid, a separate power board, standard WR-10 flanges and SMA connectors, with the RF cavity split into multiplier and driver, PA, LNA and mixer, and IF and control zones. Sequence the build: clean the housing and carriers; attach and inspect the low-speed power board; attach the RF carriers and shims; check continuity and insulation after cure; attach the RF die and critical capacitors; bond RF and DC; inspect under the microscope; fit the partition walls, feed-throughs and the IF and control board; finally fit the lid and the waveguide flange. The adhesive volume, cure profile and bond parameters for the PA and the LNA belong in separate process cards. Tighten lid screws from the centre outwards in a cross pattern at uniform torque, and take all formal performance data with the production lid fitted.

Bring-up can be simple but the gates cannot be skipped. Apply control power only and check regulators, DAC zero, temperature and fault latching; then bias the LNA and multiplier at low current and confirm the current does not jump; then open the path from low power through multiplier, driver and final stage, watching the spectrum and the current stage by stage; then inject -50 dBm continuous wave for receive alignment and go down to -70 dBm for noise and gain; only then run transmit and receive together, starting from high attenuation. Keep three lines of record for every step: before, after, and the conditions. When power is low, check input, bias, reference plane and connector repeatability before touching the match. When spurious content appears, rule out analyser image and clock coupling before blaming the die. When performance changes once the lid is fitted, look at cavity modes, probe clearance and bonds touching the lid. With transmit and receive running together, an LO pulled by the final stage points to the LO buffer, the splitter, the ground return and the supply; a compressed receive front end is fixed by physical isolation, timing and antenna isolation, not by algorithm.

The development bench needs a 67 GHz vector network analyser with WR-10 extenders, sources with multiplier heads, a W-band power meter, a spectrum analyser with a harmonic mixer, a noise figure analyser, an oscilloscope, a temperature chamber and a 6.5 digit multimeter. The production line needs a fixed-frequency composite fixture, power heads, a noise reference source, a control board test socket and one shared analyser. Calibrate the extended VNA with TRL or LRM using standards from the same lot as the product carriers, record the coupler correction for every power measurement, confirm spectrum analyser images in two independent ways, and verify the noise source against a golden unit after warm-up. For I/Q viewing, inject a single tone and expect two quadrature sinusoids; a Lissajous figure close to a circle only shows rough orthogonality and never replaces a DC to 1 GHz amplitude and group-delay sweep. An ellipse indicates amplitude imbalance or a phase shift away from 90 degrees, an off-centre circle indicates DC offset, and spikes or steps point to digital crosstalk or the supply. Use high-impedance differential or active probes, because an ordinary ground clip disturbs the low-frequency return.

Before ramping, build a 30-unit engineering batch rather than starting at a hundred. Cross two die lots, three carrier lots and two operators, and record output power, gain, noise figure, current, I/Q balance, flange return loss and case temperature for every unit. With +18 dBm typical and +17 dBm minimum, a batch mean of +18.1 dBm and a standard deviation of 0.35 dB leaves about +17.05 dBm at mean minus three sigma, which is no margin at all; raise the internal limit to +17.6 dBm and change the design instead of relying on screening. Set the noise figure internal limit at 3.6 dB and hold I/Q at plus or minus 1.5 dB and plus or minus 4 degrees at room temperature, leaving room for temperature and digital calibration. Fixtures use dowel pins, quick clamps and precision adapters to repeat the WR-10 flange, and blind-mate or spring-pin IF connectors need their low-frequency phase consistency assessed. Verify the fixture every morning against a golden unit at the centre value and another near the limit. SMT programs, bond parameters, dispense weight and torque tool calibration must all be traceable from a module serial number, and any change of adhesive, wire, board house or die lot requires an engineering assessment with first-article data and a small-batch comparison. Modules that have been opened do not ship under the sealed-unit flow.

8. Same front end, other missions

Ground-based high-resolution sensing. The classic conflict is that closer targets return more power and compress the receiver, while small targets want full transmit power and full bandwidth. Give the system room to adjust: 0 to 20 dB of digital attenuation in the transmit chain in 1 dB steps with 3 to 6 dB of routine back-off at the final stage, 6 to 12 dB of switched attenuation or a variable gain stage after the first LNA, and differential ADCs at the I/Q output. After a radome is fitted, re-check the IF zero and the leakage: radome loss and phase at 94 GHz vary with thickness, moisture and temperature, and internal reflections create false targets. Measure radome transmission and reflection peaks with a continuous-wave sweep before changing any circuit.

Instrument front ends. This mission wants frequency accuracy, traceable amplitude, remote control and long-term stability. The housing can be larger with more cooling, and the panel should expose a reference input, a trigger and status outputs. Add a step attenuator on transmit and a calibration source injection point on receive. After a 30 minute warm-up, log power and frequency every hour for 48 hours, then run power cycles. A repeatable drift against temperature calls for a look-up correction; a random drift calls for checking connectors, supplies and the source before writing any algorithm.

Multi-channel arrays. At four or eight channels the problem moves from single-path performance to channel-to-channel consistency. The LO distribution must be symmetrical, the heat from each PA changes the gain and phase of its neighbours, and pulses on the supply bus make the LNA and the phase-locked loop jitter together. Build two channels first, confirm LO isolation, thermal crosstalk and the calibration algorithm, and only then replicate to four. Below eight channels a brick-style sub-module is easier to yield than one large carrier, where a single bad site scraps the whole board at the cost of a few extra transitions and phase error. Separate the static amplitude and phase table from the dynamic temperature correction, and inject calibration through an on-board coupling network.

9. Keep the experience inside the team

Compress the whole exercise into five numbers: +20 to +21 dBm on board at the PA and +18 dBm at the flange; a 3.6 dB internal limit on receive system noise figure; no more than 0.55 dB of total loss ahead of the LNA; +9 to +12 dBm at the multiplier output with X7 and X9 held down at the same time; and final-stage dissipation measured at the back-off point rather than estimated from the Psat line. Put those five numbers into schematic nodes, layout dimensions, assembly cards and production limits, and the front end moves from lighting up to shipping reliably.

Link item Key figure and internal limit Why it matters
Receive front-end passive loss 0.25 dB transition plus 0.30 dB pre-filter and protection, 0.55 dB total Adds directly to system noise; 0.3 dB more costs 0.3 dB of margin
First-stage LNA 2.8 dB noise figure, 23 dB gain, about +1 dBm P1dB Sets cascaded noise; its gain suppresses later contributions
Receive system noise figure 3.38 dB calculated, 3.6 dB internal limit, 4.0 dB acceptance limit Reserves margin for connector repeatability and temperature
Multiplier output +9 to +12 dBm on board, flatness 1.5 dB, X7 and X9 at -35 dBc Leaves driver margin and controls harmonic spurious
Final PA +20 to +21 dBm on board, +18 dBm typical and +17 dBm minimum at the flange Keeps every unit above the output threshold at all temperatures
Final-stage dissipation 3 W DC, about 2.87 W dissipated, about 4 percent PAE in back-off Drives housing cooling and long-term gain stability
I/Q balance Plus or minus 1.5 dB amplitude and plus or minus 4 degrees phase at room temperature Sets image rejection and demodulation accuracy

Write every anomaly as one reproducible line: what changed, why, the data before and after, and the side effect. For example: "Gate voltage moved from -0.62 V to -0.58 V; output at 94 GHz rose from 17.4 to 18.1 dBm; current increased by 35 mA; noise figure unaffected; harmonics degraded by 1 dB; accepted, thermal re-test required." Three months later that record is the team knowledge base. A notebook that says "fixed it" is a notebook that says nothing.

Supply and customization

Shenzhen Huashi Digital Technology Co., Ltd. designs, manufactures and sells W-band devices and assemblies covering 75 to 110 GHz LNAs, 80 to 100 GHz x8 and x9 multipliers, 70 to 110 GHz mixers and image-reject I/Q mixers, 92 to 96 GHz GaN power amplifiers and DC to 130 GHz detectors. For systems specified at 92 to 96 GHz with +17 dBm minimum output, a 4 dB noise figure ceiling, DC to 1 GHz I/Q baseband and WR-10 with SMA interfaces, Huashi supplies die, modules and complete front-end assemblies, with customization and engineering batches. On the process side Huashi offers thin-film carrier assembly, ribbon bonding and cavity build, AuSn and high-conductivity silver adhesive window comparison, integrated tolerance control for WR-10 probes and waveguide filters, and serial-number traceability with bond pull, die shear and failure analysis support.

For part numbers, pricing and customization, contact the Sales Department at 18688706870, or visit www.szwisi.com.

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