SDR Product Handbook: choosing a software-defined receiver or RTSA

1How to use this handbook

This is a reference document about a market, not a catalog. It is arranged so that the physics comes before the products and the products come before the recommendations, because the order matters: almost every bad SDR purchase this handbook is meant to prevent came from comparing two datasheets without knowing what the numbers on them were free to mean.

Table 1 is the router. Find the row that matches why you opened this, and start where it says.

What this handbook will not do

It will not tell you that one manufacturer's radio is worse than another's. That is partly a matter of policy, since a quantitative comparison between named products is a claim somebody has to stand behind in front of a customer, and partly because it would be bad reference writing. The devices in this market are not ranked along one axis. They occupy different points on several, and a device that is wrong for one job is the obvious answer to another.

What it will do is name the axes, say where the boundaries fall and why, and cite a source for every external fact so that a reader can check it. Where a Berkeley Nucleonics product is the answer, this handbook says so and shows the arithmetic. Where it is not, it says that too, and section 45 exists for exactly that purpose.

2What software-defined actually moved

A radio has always been a chain: capture, select, amplify, convert, decide. What changed over the last three decades is not the chain but the position of one line drawn across it, the line between what is fixed in hardware at manufacture and what can be changed later in software.

In a fully analog receiver that line falls at the far right. The demodulator is a circuit, the filter is a component, and changing the modulation means changing the radio. In a software-defined receiver the line has moved left, past the demodulator, past the filtering, in some designs all the way to the antenna. Everything to the right of the line is code.

Two things pushed it there. Converters became fast enough to digitize a useful span directly, and general-purpose processors became fast enough to do the arithmetic that used to need dedicated silicon. Neither was a radio-engineering advance. The radio world inherited them.

3How the market got here, and where it is going

Three shifts made this market, and knowing them explains why its products cluster the way they do and what is likely to change next.

• Converters crossed a threshold. Once an analog-to-digital converter could sample fast enough and with enough dynamic range to digitize a useful span, the filtering and demodulation that had been circuits became arithmetic. This is what made a general-purpose receiver possible at all.
• Integration collapsed the cost. A complete receive and transmit chain on one part, with a published reference design around it, turned building a radio into an integration exercise. That is the single largest reason there are so many products and so little difference between many of them.
• The software became the asset. Open frameworks and abstraction layers accumulated a body of validated processing that no single vendor could match, which moved the durable value from the hardware to the code and made hardware portability a purchasing requirement rather than a nicety.

Two directions follow from those, and both are visible now rather than speculative.

• Interoperability is becoming a specification. The DIFI work described in section 23 exists because digitizing an intermediate frequency destroyed the natural interoperability that an analog IF had, and buyers noticed. Expect transport format to appear on requirement lists where it used to appear on nobody's.
• The interesting frequencies are above the integrated-transceiver ceiling. Millimeter wave work, satellite ground segments and modern radar all live where a single integrated part cannot reach, which is where the architectural distinction in section 5 becomes a purchasing distinction.

4Three architectures, and what each one gives up

Almost every receiver in this market is one of three things. The differences between them are not a matter of quality; each solves the conversion problem differently and pays for it somewhere else, and knowing which one you are holding explains most of its behavior.

Figure 1 understates one thing. These are not exclusive: a real instrument often uses a superheterodyne front end to reach high frequency and a zero-IF or digital down-conversion stage behind it, so the question is rarely which architecture a device is but which one is doing the work at the frequency you care about.

5The silicon underneath, and why platforms cluster

Here is the single most useful fact about this market, and the one that explains the most at the least cost. A large share of the software-defined radios on sale are built around the same family of integrated transceiver. A second cluster has grown up beside it on a different kind of silicon, and section 6 takes that one; this section takes the cluster that still accounts for most of the units.

Analog Devices publishes the AD9361 as a 2 by 2 agile transceiver whose receive local oscillator covers 70 MHz to 6.0 GHz and whose transmit oscillator covers 47 MHz to 6.0 GHz. The AD9364 is the 1 by 1 variant, and the two share a register map and a package, so a design can migrate between them. Analog Devices also publishes a complete open reference design around the family, with schematics, gateware, firmware, embedded Linux and host drivers.

Two consequences follow, and they point in opposite directions. The good one is that a huge amount of open work, driver support, framework integration and community knowledge accumulates around a common part, which is why an engineer can get a signal on screen in an afternoon. The awkward one is that a specification shared by a dozen products is not a differentiator, and a purchase decision made on it is a decision made on nothing.

6The second cluster, and where the digitizer moved

The account above is true and it is incomplete, because it describes a part that first shipped in 2013. A second cluster has formed since, and it is built on the opposite idea. Rather than mixing a signal down to baseband and digitizing it there, it digitizes the radio frequency directly and does the mixing in arithmetic.

The representative parts are radio-frequency systems-on-chip: converter tiles, programmable logic and processor cores on one die. AMD publishes the Zynq UltraScale+ RFSoC family with sampling rates reaching several gigasamples per second, its digital-front-end variant carrying the direct-sampling input range to 7.125 GHz. And this cluster is moving fast. In December 2024 AMD announced the Versal RF series, which it describes as the fifth generation of its direct-RF devices, with direct sampling quoted to 18 GHz at up to 32 gigasamples per second, a pairing worth reading with section 48 in hand, since 32 GSPS puts the first Nyquist zone at 16 GHz, so the top of that input range is deliberate higher-zone operation rather than first-zone sampling. The practical effect is that the boundary drawn in section 7 keeps moving outward: the mixer, the channelizer and the filtering that used to be analog become logic, and the frequency at which that stops being possible rises with each generation.

For a buyer the second cluster matters in two ways. It is where the wideband end of the market is going, so a requirement written today against a few tens of megahertz may be unwritable in five years. And it is the reason a development board can now quote a bandwidth figure that used to belong to instruments, which makes the question of what a number is guaranteed against more important than it was, not less.

7Where the programmable boundary falls, drawn

Draw the line between what is fixed in hardware and what is left to software, because every device in this handbook can be placed on it and the placement predicts most of its behavior.

Figure 2 also explains why the phrase software-defined radio has stopped discriminating between products. Nearly everything on sale is software-defined by some reading of the term. The useful question is not whether a device is software-defined but how far left the line is drawn and what crosses it, which is a question with a checkable answer.

8Sampling, and the three numbers people confuse

Three quantities get used interchangeably in this market and they are not the same: sample rate, instantaneous bandwidth and usable bandwidth. Getting them straight removes a large class of purchasing mistakes.

A real band-limited signal needs a sample rate of twice its bandwidth, because each real sample carries one degree of freedom and the spectrum is conjugate symmetric. A complex sample carries two, so a complex sequence at rate f_s represents a full span of width f_s about the tuned center:

No information is gained by equation (1). B complex samples per second is 2B real numbers per second either way. What changes is the bookkeeping, and almost every SDR datasheet quotes the complex form without saying so.

9Dynamic range, and why the bit count misleads

The second field people over-read is converter resolution. An ideal converter quantizing a full-scale sine into b bits achieves a signal-to-noise ratio of:

Equation (2) is an upper bound nobody reaches. Real converters fall short, and the honest figure is the effective number of bits derived from measured signal-to-noise and distortion:

Figure 3 explains a pattern buyers notice and rarely account for. Two devices quoting the same frequency range and the same bandwidth can differ by tens of decibels in what they can see next to a strong signal, and nothing in the headline specification says so. Processing gain recovers some of it, because narrowing the resolution bandwidth by a factor of k lowers the noise floor by ten times the log of k, but processing gain cannot recover a signal the converter clipped.

10The two ends of the range, and the span between them

The bit count describes the converter. What a buyer needs is the range of the whole receiver, and that is set at the bottom by noise and at the top by distortion. Four quantities fix it, and a datasheet that omits them has not described the device.

The constant in equation (4) is thermal noise power density at the 290 K reference temperature, which is a convention rather than a law: a cryogenically cooled front end sits below it, and a warm antenna looking at a warm scene sits above it. For any receiver operating at room temperature it is the practical floor, so a sensitivity claim can be checked against it in one line. A receiver with a 10 dB noise figure measured in a 1 kHz bandwidth cannot hear below −174 + 10 + 30, which is −134 dBm, and a claim of −150 dBm in that bandwidth is not a good product but an arithmetic error or an unstated bandwidth.

Table 3 is the set of four.

11Phase noise, and the signal you cannot see next to a strong one

Noise and distortion set the two ends of the range. A third mechanism decides what you can see next to something strong, and it is the one most often left out of a comparison because it does not appear as a single number.

No oscillator produces one frequency. A real local oscillator produces a carrier with skirts of noise either side, and phase noise is the power in those skirts, quoted as decibels relative to the carrier in a one hertz bandwidth at a stated offset. So a figure like −110 dBc/Hz means nothing until you are told the offset it applies at, and a vendor quoting one offset has chosen the flattering one.

Two consequences follow that a datasheet comparison will not show you. A receiver with an excellent noise figure and a mediocre oscillator finds weak signals in a quiet band and loses them next to a strong neighbor, which is the situation every real deployment is in. And an adjacent-channel measurement is a measurement of your instrument's oscillator until you prove otherwise, which is why the honest way to report one is with the instrument's phase noise at the offset in question stated alongside it.

12What zero IF costs, and where it shows

The zero-IF architecture mixes the wanted band straight down to baseband as an in-phase and a quadrature stream. It is elegant, it is cheap, it integrates well, and it has three characteristic impairments that every user of such a device eventually meets. They are worth naming because they are visible on a screen, and once you can name them you stop misdiagnosing them as signals.

• A DC offset at the center of the span. Any offset in the baseband path appears at zero frequency, which after conversion sits exactly at the tuned center. It looks like a carrier and it is not one. Tuning deliberately off the signal of interest, which is what a local oscillator offset does, moves the wanted signal away from it.
• Quadrature imbalance, seen as an image. If the I and Q paths differ in gain or are not exactly ninety degrees apart, a signal at one offset produces a mirror at the negative of that offset. The mirror is an artifact of the receiver and moves when you retune, which is the test that distinguishes it from a real emitter.
• Local oscillator leakage. The oscillator is at the tuned frequency by construction, so any leakage back to the input reappears at the center of the span, reinforcing the same spot the DC offset occupies.

13The five settings that decide what a screen shows

Every instrument capture in this handbook was taken with five controls set, and every one of them changes the answer. A reader who has only ever used a receiver through a flowgraph will not have met them, and a reader who has used an analyzer for years may never have been told what they do. They are the difference between two people pointing the same instrument at the same signal and reporting different numbers.

Table 4 is the set.

14Processing gain, and what it can and cannot recover

One more relation belongs in Part I because it is used to justify purchases it does not support. Narrowing the resolution bandwidth lowers the displayed noise floor:

Equation (5) is why a signal invisible in a 1 MHz resolution bandwidth appears cleanly in 1 kHz: a factor of a thousand is 30 dB of noise floor. It is real, it is free, and it is the reason a narrow-resolution sweep finds things a wide one does not.

That second point is the trade of section 33 arriving from the other direction, and it is worth stating as a rule. Sensitivity and intercept pull against each other through the same product, so a specification that demands both an extremely low noise floor and a microsecond intercept guarantee is asking for two settings at once. Decide which the application actually needs, or specify two configurations and accept switching between them.

15Noise, gain and the front end you attach

A receiver's published noise figure describes the receiver. What decides whether you hear anything is the whole chain from antenna to converter, and the first stage dominates it. The cascade relation settles most front-end arguments on its own:

Worked once, the relation stops being an argument. Take a receiver with a 15 dB noise figure and three decibels of cable ahead of it. A passive loss contributes its own loss as noise figure, so the chain comes to 18.0 dB and the cable cost exactly what it lost. Now put a 1.0 dB, 20 dB-gain preamplifier at the antenna, ahead of the same cable and the same receiver. The relation gives 1.259 plus 0.010 plus 0.611, which is 1.880, and ten times its log is 2.74 dB. The receiver did not change and the chain improved by 15.3 dB, all of it bought by twenty decibels of gain standing in front of everything else. The bill arrives on the other axis: every signal reaching the receiver is now 20 dB stronger, so the compression and intercept points of the whole chain drop by about the same 20 dB, and a preamplifier that rescues a quiet rural site will drive the same receiver into distortion within sight of a broadcast mast.

Three practical consequences follow, and they apply whatever class of device you buy.

• Cable loss before the first amplifier is noise figure, decibel for decibel. Several decibels per meter is normal in flexible assemblies at millimeter-wave frequencies, so a long run to a mast-top antenna can cost more sensitivity than any receiver choice will recover. Amplify at the antenna or accept the loss knowingly.
• Choose the antenna before the receiver. A broadband discone gives up roughly 6 dB against a log-periodic and considerably more against a horn on the same band, and the antenna factor has to be entered before any field-strength number means anything at all.
• Preselection is what lets a wideband receiver work in a real environment. A device that digitizes a wide span sees every emitter in that span at once, so the strongest one sets the gain and everything weak lives in whatever dynamic range is left.

16Reading an SDR datasheet

An SDR datasheet is a shorter document than an instrument datasheet, and the difference is not laziness. Most of the fields an instrument publishes describe accuracy, and a device that makes no accuracy claim has nothing to put in them. Knowing which fields are missing is more useful than reading the ones that are there.

Table 5 is the field list, with what each one is worth.

17The classes of device

Products in this market fall into four groups. The grouping is by where the boundary between fixed hardware and user responsibility falls, which is the distinction from Part I applied commercially.

Named examples make the classes concrete, so here are the ones a newcomer actually meets. The first row is where most people start: receivers built on the RTL2832U demodulator, sold as RTL-SDR sticks, and the Airspy and SDRplay RSP families a step above them. The second row is the research bench: the ADALM-PLUTO, the USRP B-series, LimeSDR and bladeRF, most of them built on the transceiver family of section 5, alongside the HackRF as a wideband half-duplex outlier. Evaluation boards for the direct-sampling parts of section 6 sit at the top of this row at a very different price. The third row is rarely famous, because a module that ends up inside somebody else's product is sold on supply life and interface stability rather than on brand. The fourth row is where the ICX line sits, and it is a different category rather than a better second row.

These are named as landmarks, not as recommendations, and nothing here compares them against each other or against ours. Each publishes its own specifications; read them against the field list in section 16 and the class becomes obvious from what is missing.

Read Figure 4 in both directions. Buying up the stack when the job did not need it wastes money on accuracy nobody will use. Buying down it, which is far more common, produces a result that cannot be defended when somebody asks how it was calibrated, and the cost lands months later when the campaign has to be repeated.

One more thing follows from Figure 4 that is easy to miss. The classes are not a price ladder with the same product at four qualities. They differ in what the vendor is prepared to be held to, which is why a device three rows down can outperform one two rows up on a headline number and still be the wrong purchase. Buy the row, then the model.

18Transmit, duplex and multiple channels

Three capability questions separate otherwise similar devices, and each has a trap in it.

• Transmit is not the mirror of receive. A device that receives well may transmit poorly, and transmit specifications are usually thinner: an output power figure at one frequency, and often nothing about linearity or spurious content. If the transmitted signal has to be clean rather than merely present, that is a question to ask before purchase.
• Half duplex and full duplex are different products. A half-duplex device transmits or receives, not both, which rules out any application that has to hear itself or hear a reply during transmission. The datasheet word to look for is the one that is often missing.
• Two receivers are not a coherent pair. Two channels that happen to be on one board are not phase coherent unless they share a reference and their relative phase is specified. Direction finding, beamforming and time-difference work all need the second property, and it is the one least often published.

19Antennas, preselection and the part nobody budgets for

A receiver is half a system. The other half is what is in front of it, and it is routinely specified last, bought cheapest and blamed least when the numbers disappoint.

Table 6 is the antenna decision in the order it actually matters.

Preselection is the other half of the same subject. A wideband receiver sees every emitter in its span at once, so the strongest one sets the gain and everything weak lives in what is left. In a broadcast-heavy environment a filter ahead of the receiver can buy more usable dynamic range than any change of receiver, and it is cheaper. The trade is that a filter you install is a band you can no longer see, which is why the question belongs in the requirement rather than in the field.

20What each class actually costs you to own

Table 7 is the part of a purchase that no datasheet covers: what the device asks of you after it arrives. The rows are the classes from Figure 4, and the columns are the costs that show up in a project plan rather than on an invoice.

21Timing, reference and synchronization

Timing is where distributed systems succeed or fail, and it is consistently the weakest section of an SDR datasheet. Three quantities matter and they are frequently confused.

Table 8 separates them, because a datasheet that answers one is often read as having answered all three.

Memorize one conversion, because it turns a timing specification into a distance and makes the consequences obvious. Light travels 0.2998 m in a nanosecond, so an arrival-time uncertainty of a hundred nanoseconds is thirty meters of range uncertainty before any geometry is applied. BNC-AN-110 works that arithmetic through for a distributed radio-location system and finds that the receiver's own contribution is not the term that decides the answer.

22Form factor, environment and the things that break

A receiver spends its life somewhere, and where decides more about the purchase than most specification sheets acknowledge. Three questions separate a device that works in a laboratory from one that works where the signal is.

• Power and thermal. A module drawing ten to twenty watts inside a sealed enclosure is a thermal problem before it is an electrical one, and enclosure temperature moves amplitude through the receiver's temperature coefficient. Published operating classes matter: a base class covering 0 to 50 °C is a laboratory instrument, and a class reaching −40 to +65 °C is a different product even if the RF specification is identical.
• Ingress and mechanical protection. A sealed chassis is the difference between a device that survives a wet mast and one that survives a demonstration. Read the specification row rather than the prose: ingress ratings are frequently described in marketing copy and absent from the table.
• Damage limits. Every receiver has a maximum input beyond which something stops being reversible. A device with a preamplifier switched in usually has a much lower limit than the same device without, and a field operator who forgets that is the commonest cause of a receiver going back to the factory.

23Interfaces, transport and the standards that matter

How samples leave the device is a system decision disguised as a connector choice.

• USB is convenient, ubiquitous and shares its bandwidth with everything else on the controller. It is the reason a great many overflow problems are diagnosed as software faults when they are transport faults.
• Ethernet decouples the receiver from the host and lets it sit where the antenna is, at the cost of needing the network to behave. For distributed sensing this is usually the right answer.
• PCIe gives the most throughput and the least flexibility about where the receiver is.

Above the connector there is a question of format, and this is where the market has done something genuinely useful. VITA 49 is an ANSI standard for carrying digitized RF with its metadata, widely used in satellite and software-defined radio work. Its flexibility is also its problem: two compliant implementations can still fail to interoperate because the standard defines what is possible rather than what is required.

The DIFI Consortium was formed under the IEEE Industry Standards and Technology Organization to fix exactly that. Its Digital IF Interoperability Standard, published as IEEE-ISTO Std 4900-2021, is a constrained profile of VITA 49.2: a specific schema that removes the ambiguity, so that where VITA 49 defines what is possible, DIFI defines what is required. The consortium's stated purpose is to replace the natural interoperability that analog intermediate-frequency signals used to have, and to prevent vendor lock-in.

One practical note on asking. The standard has been revised since its 2021 designation, so a supplier who answers "we support DIFI" has not answered the question. Ask which version, and ask whether they have run the consortium's certification material against it, because a packet format is only interoperable to the extent that two implementations have actually been made to talk to each other.

24Frameworks, abstractions and applications

Software for this market divides into three layers you choose between, resting on a fourth you do not, and confusing them is the source of most bad architecture decisions. A framework is where you build signal processing. An abstraction layer is how any framework reaches any radio. An application is a finished tool somebody else built.

The names are better stated than talked around, because a reader cannot evaluate an architecture described only in categories. GNU Radio is the dominant open framework: a scheduler, a block library and a graphical flowgraph editor, with a long history and an annual conference. SoapySDR is the dominant open abstraction layer: a small C and C++ API with language bindings, against which a driver module is written per device family, so a flowgraph written for one radio runs on another. Around them sit finished applications: general-coverage receivers such as GQRX, SDR# and SDRangel, and single-purpose decoders such as dump1090 for aircraft transponders, which are what most people meet first. Vendor software development kits occupy the same layer as the abstraction layer, with the opposite trade: everything the hardware can do, for exactly one vendor.

Figure 5 carries a warning worth stating directly. The bottom row is the only place calibration can come from. An application cannot add it, a framework cannot add it, and an abstraction layer explicitly does not carry it. If the hardware does not produce corrected samples, no amount of software above it will produce a calibrated measurement.

25Abstraction layers, and what they deliberately omit

The third layer of the software stack is the one that decides whether the work in the second survives a change of hardware, and it is chosen with the least deliberation of the three. It deserves more, because the decision is close to irreversible once a body of processing has been written against it.

The problem it solves is narrow and expensive. Every vendor ships a programming interface, and each answers the same first questions differently: how a device is found, whether the tuner is one frequency or a chain of elements, whether gain is a scalar or a set of named stages, what format samples arrive in, who owns the buffer. Application code inherits all of those answers. The demodulator and the decoder are portable and expensive to validate; the forty lines that open the radio are neither, and they are what make the rest device-specific.

Table 9 sets out the four ways a system can reach its hardware.

None of that is a defect. An abstraction that tried to standardize calibration across hardware that mostly has none would not have been adopted, and adoption is the whole value of an abstraction layer. But it does mean the question that decides whether an open framework can produce a defensible number is a question the abstraction cannot answer, and the buyer has to put it to the hardware instead. BNC-HB-001 documents one such interface in working detail, including the call that returns full scale and why it is the one the whole argument turns on.

26The arithmetic that decides whether a system is buildable

Two calculations settle most architecture arguments before any hardware is chosen, and both are short enough to do in a meeting.

First, the data rate out of a receiver:

At 125 MSPS with 16-bit components, equation (7) gives 500 Mbyte per second. That is more than a mechanical disk will take and about what a SATA solid-state drive will sustain, which is why a wideband capture architecture is a storage design as much as a radio design. An NVMe drive has several times that headroom, but it buys the headroom with power and heat, which is exactly the budget a battery-powered field instrument does not have spare. The number that matters is the sustained rate, not the burst rate, and section 16 explains why a device quoting only one figure has answered the easier question.

Keep Figure 6 in view during any architecture discussion, because the storage question is usually raised last and decides the design.

Second, how long a buffer holds. A capture memory divided by the rate above gives the window you can take away and reprocess. For the ICX-FieldHawk's published 128 Mbyte of capture memory at 500 Mbyte per second that is about 0.26 seconds of gap-free 100 MHz capture, assuming a 16-bit component width, which is a conventional interleaved format rather than a published property of the instrument.

27How a processing framework actually runs

The framework layer is where custom work lives, so it repays understanding one level below the block diagram. BNC-HB-001 documents this in depth; what follows is the part that changes a purchasing decision.

A graph-based framework connects processing blocks with buffered edges. Each block runs on its own thread, asks how many items are available and how much output space is free, and does work when both allow. Every output port gets one circular buffer, allocated when the graph starts and never resized. That single fact explains the behavior that surprises people most.

Figure 7 is the reason a device's transport and a host's performance are part of the measurement chain rather than logistics. The symptom is intermittent decode failure with a perfectly healthy-looking spectrum, and the cure is almost never a faster computer: it is decimating at the source so the samples are never produced, cutting display update rates, and taking per-sample work out of interpreted code.

28Samples, scaling and time: the three that break integrations

An engineer told to integrate a receiver usually loses the first week to one of three things, and none of them appears in a datasheet. They are worth stating plainly because each has a symptom that looks like something else.

The format. Samples arrive as interleaved pairs, one in-phase value then one quadrature value, repeating. Two things about that sentence have to be confirmed rather than assumed: the order of the pair, and the type of each value. Sixteen-bit signed integers and thirty-two-bit floats are both common, and so is the byte order question, which does not arise until the day the capture host and the analysis host are different machines. The symptom of a swapped pair is a spectrum mirrored about the center, which is easy to mistake for an image problem in the radio.

The time base. A timestamp is almost never attached to every sample; it is attached to the first sample of a buffer, and the time of every later sample is inferred by counting at the sample rate. That inference is what makes timing cheap, and it is also what makes a dropped sample dangerous. A lost sample does not leave a hole. It shifts everything after it earlier by one sample period and leaves no evidence in the samples themselves. The interface may still tell you, which is the whole point of asking whether it does. In a 125 MSPS stream one lost sample is 8 ns, which is nothing; a hundred thousand lost samples during one disk stall is 0.8 ms, which is more than enough to break a time-difference measurement or a correlation between two units.

None of these three is difficult once named. All three are expensive when they are not, because each produces a plausible wrong answer rather than an obvious failure, and a plausible wrong answer can survive a long way into a project.

29Platforms, and where RF data goes after the disk

Two practical questions close out the software part, and both are usually discovered after a purchase rather than before it.

What runs it. A receiver's usefulness is bounded by the machines its software supports. The ICX line publishes Windows 11, 10, 8 and 7, Debian 12, 11 and 10, and Ubuntu 24.04 back to 18.04, on x64 and AArch64. The AArch64 entry is the one worth noticing: it means a single-board computer can host the receiver, which is what makes an unattended node affordable, and it is the row most often missing from a competing datasheet. Ask for the list, and ask how long each entry is supported, because an operating system that falls out of support takes your deployment with it.

Where the data goes. Section 26 prices the disk; it stops there, and most systems do not. Three patterns are common and they have different failure modes. Keeping everything centrally is simplest and is bounded by the link, not the disk: a single sensor streaming continuously will exceed most site uplinks, so the architecture becomes a conversation about bandwidth rather than storage. Processing at the node and sending results is what makes a network of sensors affordable, and it moves the hard problem to deciding what a result is, because whatever the node discards cannot be reanalyzed when the question changes. Sending triggered captures is the middle path and the most common answer, which puts the weight back on section 34: the trigger becomes the compression algorithm, and its settings become part of the evidence.

30Diagnosing an overrun, and testing a pipeline

Two pieces of software practice are worth more than any amount of architecture advice, because they are what the first month actually consists of.

Triage an overrun in this order, because each step is cheaper than the one after it and rules out a whole class.

• Confirm it is real. Read the stream return value, not the console. A pipeline that reports nothing may be losing samples silently, which is the case section 28 warns about.
• Ask whether it happens immediately or under load. Immediately means a configuration error, usually a rate the host was never going to sustain. After minutes means a resource that fills or degrades: a disk buffer, a log, a thermal throttle.
• Time the blocks individually. One slow block stalls everything upstream of it through the back pressure of section 27, and the symptom appears at the source rather than at the culprit. Replace the suspect with a null sink and see whether the overruns stop.
• Take the disk out of the loop. Write to a null device at the same rate. If the overruns vanish, the problem is storage, and the fix is the arithmetic of section 26 rather than anything in the radio.
• Then look at the host. Power management that drops clock speed, an unpinned thread migrating between cores, memory that is not locked, and an interface with a shallow buffer are the usual remainder. A machine that idles at low clock speed will overrun a stream it can easily sustain at full speed, and that one accounts for a surprising share of pipelines that worked yesterday.

One representation detail belongs here because it is invisible until it is expensive. A timestamp carried as a floating-point count of seconds since the epoch cannot hold nanoseconds. A double has about sixteen significant digits, and at a present-day epoch value that leaves about 238 nanoseconds of resolution, which is thirty samples at 125 MSPS. For a spectrum display that is irrelevant. For time-difference work between two units it is the entire measurement. Carry integer seconds with an integer fraction, or a 64-bit count of nanoseconds, which spans a couple of centuries and loses nothing.

31What a software lead should ask before anyone commits

The person who will carry the integration is usually not in the room when the receiver is chosen, and the questions that decide their year are not on any datasheet. Six are worth asking early enough to matter.

• What language is the fast path in, and where does the interpreter sit? A graph assembled in Python whose blocks are compiled is fine, because the interpreter only builds the graph. A custom block written in Python sits in the sample path and takes the interpreter's global lock with it, so it serializes against every other Python block in the process. This is the most common reason a pipeline that worked on a bench recording will not keep up with a live receiver.
• How many times is a sample copied between the converter and the first useful operation? Driver to library, and library into the framework's buffer, are usually copies. What happens after that often is not: a well-built framework hands a block a pointer into its upstream's output buffer rather than a copy, which is why the buffer model in section 27 is worth understanding. The copies that remain are the ones you pay for, and at 500 Mbyte per second memory bandwidth is a real budget rather than a rounding error.
• What is the driver's dependency and version story? An abstraction layer loads modules built against a particular version of itself, so a distribution upgrade can remove a working radio from a working machine without anyone touching the radio. Ask which versions are supported, how long, and whether a build from source is a supported path or a rescue operation.
• How do we test this without a signal generator on every desk? A pipeline that can be fed from a recorded file is testable in continuous integration; one that can only be fed from hardware is testable when somebody is in the building. Ask for a recorded capture with known contents, and treat it as a deliverable rather than a favor.
• What is the support window, and how are updates delivered and signed? The Cyber Resilience Act of section 54 turns this from a courtesy into a regulated commitment, and it is the question that decides whether the device is still deployable in five years.
• What is the calibration, concretely, on disk? If correction is applied before the samples reach you, then it is a file that was installed alongside the driver, it is specific to that unit, and it has a date on it. Ask where it lives, what happens if it is missing, whether the software warns or silently continues, and what the update path is after a recalibration. A correction that can go missing quietly is a correction you cannot rely on, and this question is rarely asked until the first time it happens.

32What separates a reading from a measurement

A number without a stated uncertainty and a traceable reference is a reading. A number with both is a measurement. Everything in Parts I to III is about producing readings well. This part is about the other thing, and the distinction is the single most consequential one in the handbook.

An uncorrected receiver reports its converter's output as a fraction of full scale. Turning that into an absolute power needs a correction table covering every gain state and every frequency in use, built against a reference source, maintained as it drifts and re-checked when anything in the chain changes. Nothing prevents an engineer from building that table. It is a large, dull, recurring job that most projects discover late.

This is also why the absence of an amplitude accuracy figure on a datasheet matters more than it looks. A receiver quoting plus or minus 2.0 dB is making a claim about every reading it produces under stated conditions. A board quoting nothing is not making a weaker claim; it is making no claim, and the difference between those two positions is the difference between a result that can be published and one that cannot.

33Real-time capture, and what gap-free guarantees

The second dividing line is what happens to signals that are not on all the time. A swept receiver visits each frequency in turn, and anything happening elsewhere while it looks here is absent from the result. The absence is silent: the trace looks the same whether the band was quiet or the receiver was pointed the other way.

A real-time receiver holds a span continuously and transforms every sample. Where the governing relations are published rather than only their results, a reader can price any configuration:

The factor of two in equation (8) is not margin. A burst exactly one frame long, arriving halfway through a frame, puts half its energy in each of two frames and is measured at full amplitude in neither, so guaranteeing full amplitude means guaranteeing that a whole frame falls inside the burst.

That coefficient belongs to contiguous, non-overlapping frames, which is the arrangement a relation of this form describes. An engine that overlaps its transforms needs less: where a new frame begins every half frame, any burst lasting one and a half frame lengths already contains a whole one. So the leading coefficient is part of the relation a vendor has to publish rather than a constant of nature, and a competitor quoting 1.5 is not necessarily flattering the number. Asking which coefficient applies is the second question, after asking for the relation at all.

Table 10 evaluates it at five settings of the ICX-FieldHawk real-time engine.

The last row is the end of the published range rather than an arbitrary stopping point, and the reasoning shows how to check a datasheet against itself. The published real-time resolution bandwidth runs 7.81 MHz down to 1.90 kHz in 13 grades under this window. The grid behind those grades is the thirteen powers of two in N D from 32 to 131,072, whose bin spacings at this engine are 3.90625 MHz down to 953.674 Hz. The column that confirms the grid is the flat-top one published beside it: 14.73 MHz divided by 3.90625 MHz is 3.771, and 3.59 kHz divided by 953.674 Hz is 3.764, both within 0.2 percent of the flat-top window's equivalent noise bandwidth of about 3.77. So 2.097 ms is the longest guaranteed intercept the published range reaches, over the 1.95 MHz span that decimation leaves, and a requirement written past it is a requirement for a different instrument. Write both numbers into any specification that uses it, because an intercept guarantee without the span it applies over is the same half-answer this handbook keeps warning about.

The first row is the purchase criterion at its sharpest. At 0.512 microseconds of guaranteed intercept, every event in the band lasting longer than about half a microsecond is measured at its true amplitude rather than glimpsed. A swept receiver at the same span offers no such boundary at all, and BNC-AN-102 works through how far apart the two answers land for a monitoring system.

34Triggering, and how a finite buffer sees a long day

The arithmetic of section 26 says a wideband receiver fills any disk you own. The way out is not a bigger disk; it is deciding what to keep, and that decision is triggering. It is the least specified capability in this market and one of the most consequential.

Table 11 is the set of modes worth knowing by name.

35Distributed systems, and what changes when there is more than one

A single receiver has one clock, one reference and one place. A network of them has as many of each as it has nodes, and almost every hard problem in distributed sensing comes from that multiplication rather than from the radios.

Table 12 sets out what changes.

36How a swept analyzer works, and the relation that governs it

The antagonist of this part deserves a fair description, because it is the right answer to a great many problems and because its limitation follows from a relation as clean as the one on the other side.

A swept-tuned analyzer moves a narrow filter across the span and records what comes through it. It sees one resolution bandwidth at a time, and the filter needs time to settle: a filter of bandwidth RBW cannot deliver a settled output faster than about 1/RBW. Dwelling that long on each of Span/RBW positions gives the sweep time.

One qualification, because it is the objection this argument attracts. The relation above is the analog swept-filter law. An analyzer that steps a transform across the span instead needs a record length of about 1/RBW per segment, with the segment count set by its analysis bandwidth rather than by its resolution bandwidth, so its sweep time falls to roughly Span / (B × RBW), where B is the analysis bandwidth of each segment: linear in 1/RBW rather than quadratic, and often far faster than the swept-filter law predicts. The contrast is not subtle. A 40 GHz span at 1 kHz resolution is about 80,000 seconds under the swept-filter law and about 0.4 seconds for a stepped transform with a 100 MHz analysis bandwidth. What no stepped architecture escapes is the second half. While it is looking at one segment it is not looking at the others, so the argument about what fraction of the time the instrument is pointed where the event is survives intact even where the speed argument does not.

Set that beside the real-time relation and the two architectures separate cleanly. A swept instrument's cost goes as the square of the resolution you ask for in a swept-filter design and linearly in a stepped-transform one, and either way it is blind to everything outside its filter while it sweeps. A real-time engine holds the whole span continuously, and its cost is a fixed intercept guarantee that goes linearly in transform size. Neither is better; they are priced in different currencies. A swept analyzer is the right instrument for a stable signal you can dwell on, for the best available resolution and dynamic range at a price, and for spans wider than any real-time engine digitizes at once. It is the wrong instrument for anything that is not there when you arrive.

37The evidence a swept trace cannot give you

Section 33 argued the point from a relation. This section argues it from a screen, because the difference between a receiver that holds a band and one that visits it is visible without reading a single number.

Figure 8 is the argument of Part IV in one picture, and it generalizes past this band. Any environment with intermittent or hopping emitters produces the same divergence, and the size of it is set by duty cycle rather than by anything in the receiver's specification. A single-sweep trace of a bursty environment is not a measurement of that environment; it is a sample of it at an instant nobody chose. The spectrogram above the traces is the same information in time, which is often the faster way to see that an emitter is present at low duty cycle rather than absent.

38What the open world does with a calibrated front end

The two halves of this handbook meet here. Part III described an open software ecosystem that can decode almost anything; Part IV described a class of hardware that produces numbers you can defend. The interesting question is what happens when they are put together, and it has been demonstrated end to end rather than argued.

Figure 9 and Figure 10 are the argument of this handbook in two screenshots. Both were produced by open blocks that were never written for this instrument, attached at a standard boundary, with the amplitude correction already applied beneath them, a behavior demonstrated in the integration guide these captures come from rather than specified on a datasheet. Neither half gave anything up to work with the other, and that is the unusual position: most of this market is a choice between open and programmable hardware that is not calibrated, and calibrated instruments that will not let you inside.

39Calibration, and what traceability actually costs someone

Part IV keeps returning to calibration, so here is what the word covers, because it is used loosely enough to have almost stopped discriminating.

• A correction is a number applied to a reading. Anyone can produce one, and a correction with no provenance improves precision and does nothing for accuracy.
• A calibration is a correction established by comparison against a reference of known accuracy, at a known date, under known conditions. It has an owner and an expiry.
• Traceability is the unbroken chain from that reference to a national or international standard, with a stated uncertainty at every step. It is what lets a number be defended by somebody who was not present when it was taken.

The chain is the expensive part, and it is why calibration has resisted commoditization while everything around it has become software. Somebody has to own reference equipment, have it calibrated against something better on a schedule, maintain an environment in which the comparison means something, and keep records that survive an audit. None of that scales down, which is precisely why a device class that does not carry it exists and is the right answer for most jobs.

Figure 11 also answers a question buyers ask in good faith: why not calibrate a cheap receiver yourself? You can, and for a single fixed configuration it is a reasonable afternoon's work. What does not scale is the second unit, the second frequency, the second gain state, and the drift that starts the day after. The correction table is not the hard part. Maintaining it, across every state the device can be in, for as long as the program runs, is the hard part, and that recurring cost is what a calibrated instrument is actually selling.

40Machine learning on spectrum, and what it inherits

A growing share of what people want from a wideband receiver is not a display but a training set. Modulation recognition, anomaly detection over a monitored band, and radio frequency fingerprinting, which identifies an individual transmitter from the small imperfections of its hardware, are all now ordinary asks. The handbook has already supplied everything needed to see where they go wrong.

This is the handbook's argument arriving in a new place. A model is a measurement instrument built out of data, and it inherits every property of the data it was built from. If the amplitude axis of a training capture has no absolute meaning, the model has learned relative structure and cannot transfer to a receiver whose gain differs. If the captures are not gap-free, the model has learned from a record with holes in it that are invisible in the file. If the time base drifts, a model that uses timing features has learned the drift.

Four practices follow, and none of them is about the model.

• Record raw samples, not processed traces. A spectrogram image discards phase and throws away most of what a fingerprinting model uses. Keep the samples and the metadata that makes them interpretable.
• Train across receivers, not just across transmitters. If every capture comes from one unit, the model cannot distinguish the transmitter from the receiver, and no amount of held-out test data will reveal it, because the held-out data has the same receiver in it.
• Calibrate before you collect, not after you fail. A correction applied to the samples makes the amplitude axis comparable across units, days and temperatures, which is what makes a dataset reusable rather than disposable.
• Keep the provenance with the data, in a format somebody else can read. Receiver, serial number, gain settings, sample rate, antenna, date, temperature, calibration due date. The field has a convention for this, the Signal Metadata Format, which pairs a binary sample file with a JSON sidecar describing it; using it costs nothing and is the difference between a dataset and a directory. A capture without provenance is not a measurement and it is not a training example either.

None of this argues for or against any particular instrument. It argues that the dataset is the deliverable, and the properties that make a measurement defensible are the same ones that make a dataset transfer.

41Where the ICX-FieldHawk sits

The ICX-FieldHawk is a real-time spectrum analyzer family built around one analysis engine and one host interface, SpectraCore, in three physical forms. It belongs in the fourth row of Figure 4, and this section says what that means in published numbers rather than in adjectives.

Table 13 is the family.

Table 14 is the performance, each row with the condition that makes it true.

42What an instrument does that a receiver does not

The abstract difference is calibration. The concrete difference is that an instrument carries measurement functions whose results are defined, repeatable and checkable from the screen. That last property is worth more than it sounds, and it is the practical test to apply to any device claiming instrument status.

Figure 12 is the property to look for, stated generally: a result you cannot recompute is a result you cannot defend. A device that reports a derived quantity and also reports the raw quantities it derived it from lets a reader check the derivation without trusting the device. That is a design choice rather than an accident, and it is rare below the instrument class because nothing else in the stack is obliged to make it.

Those functions run under SpectraCore, the host software shared across every form factor in the family, and the published list behind that behavior is long and, more usefully, is included rather than licensed piece by piece: channel power, occupied bandwidth, x dB bandwidth, adjacent channel power ratio, third-order intermodulation, spectrum emission mask, AM and FM demodulation, antenna factor, amplitude offset, signal track, peak table, data record and playback, multiple unit display and amplitude correction. The operating modes are standard spectrum analysis, IQ streaming with digital down-conversion, power detection at 8 ns time resolution, real-time spectrum, phase noise measurement from 1 Hz to 10 MHz offsets, and harmonics analysis. Digital demodulation and pulse detection are options 71 and 72.

43Four more measurements, each checkable from its own screen

The ACPR screen made a general point. These four make it four more times, and together they are the practical answer to what an instrument buys over a receiver: not prettier traces, but derived quantities you can verify without trusting the instrument that produced them.

Figure 13, Figure 14, Figure 15 and Figure 16 have one property in common that is worth naming as a purchasing test. Each reports a derived quantity and the raw quantities it was derived from, so a reader can check the derivation without trusting the device. Apply that test to any instrument you are evaluating: ask it for a measurement, then ask whether the screen contains enough to reproduce it. A device that reports only conclusions is asking for a kind of trust that a defensible measurement cannot be built on.

44One engine, three form factors

A point from Part II applies to this family specifically, and it can be shown rather than asserted: within a product line, the form factor should change what the instrument survives and where it goes, not what it measures.

Figure 17 makes the commercial case the way this handbook prefers to make it, which is by naming a property a buyer can test. Ask any vendor offering a family whether a routine written against one member runs unmodified against the others, and what specifically differs. A family whose members need different code is three products sharing a brand, and the difference shows up as integration cost long after the purchase.

The published physical differences are real and are the point of having three. Module weights start below 305 g for the 9.5 GHz USB variant and reach 665 g for the LAN version, against 1.5 kg for the handheld with its display and battery. Environmental classes are options rather than variants: a base class covering 0 to 50 °C, option 40 widening it to −20 to +65 °C and option 41 to −40 to +65 °C, published on the rugged and module datasheets.

45Where this is not the answer

A handbook that only lists capability is a brochure. These are the cases where an ICX-FieldHawk is the wrong purchase, and naming them is how the rest of the document earns its credibility.

• You need transmit. This is a receiver family. Vector signal generation is not an ICX capability; the only published generation feature is a built-in signal generator covering 100 kHz to 6.3 GHz on USB models as option 02. A record-and-replay bench pairs an ICX capture with a separate generator, and BNC-AN-105 keeps the two halves apart throughout.
• You need continuous wideband recording. Sustained capture is published at 25 MHz, not 100 MHz. If the application genuinely needs minutes of unbroken 100 MHz recording, the right answer is a recorder with a storage array behind it.
• You need DIFI or VITA 49 transport. Not published. See section 23.
• You are learning, teaching or prototyping on a budget, below 6 GHz. The development class exists for exactly this, it is excellent at it, and an instrument bought for a job that never needed traceability is money spent on a column nobody will read.
• You need a coherent array today. Multiple-unit operation is a published function, but if the requirement is phase-coherent direction finding, ask what is specified about the relationship between units before committing, exactly as section 18 advises asking any vendor.

46An uncertainty budget, worked

A bound like plus or minus 2.0 dB is the instrument's contribution. It is not the uncertainty of your measurement, and the difference between those two things is where most disputes about a number begin. Building the budget is short arithmetic and it is worth doing once so the shape is familiar.

Each contribution enters as a variance. A bound quoted as plus or minus a with no distribution stated is treated as rectangular, which contributes:

Sum the variances, take the root, and multiply by a coverage factor of two for approximately 95 percent confidence. Table 15 does that for a field measurement of the kind BNC-AN-107 works through in full.

47Bandwidth, and the four numbers a span can mean

Section 8 separated sample rate from bandwidth. In practice a span quoted on a screen or a datasheet can mean any of four things, and confusing them accounts for a surprising share of the disagreements between a measurement and a specification.

Table 16 separates them.

48Nyquist zones, and deliberate undersampling

Sampling at rate f_s divides the frequency axis into zones f_s/2 wide. The first runs from zero to f_s/2, the second from f_s/2 to f_s, and so on. Everything in every zone lands on the same output band. That is the fact aliasing is usually taught as a hazard.

Used deliberately it is a technique rather than a hazard. If the analog front end passes only one zone, sampling can be slower than the signal's frequency and still recover it completely, because the ambiguity the theorem warns about has been removed by the filter instead of by the sample rate. This is why a converter quoting an analog input bandwidth far above its sample rate is not contradicting itself. Two details matter in practice: even-numbered zones come out frequency-reversed, so the spectrum is a mirror image and any downstream processing has to know it, and the higher the zone the more the converter's aperture jitter costs you, so sensitivity falls off with zone number even though the arithmetic does not.

49Aliasing, and the spurs that are not there

One more artifact class deserves naming, because it produces the most confident wrong conclusions. A sampled system cannot distinguish a signal at frequency f from one at f plus any multiple of the sample rate. Everything outside the digitized span folds into it, and once folded it is indistinguishable from a signal that was really there.

That is what the anti-alias filter prevents, and it is why the bandwidth control of section 16 is not a convenience. Set a sample rate of 20 MSPS and leave the filter at 56 MHz and everything in that 56 MHz appears somewhere in your 20 MHz of data, at frequencies bearing no relation to where it actually was.

The third test is worth a sentence more, because it is the one that catches a receiver being driven too hard. Intermodulation products are generated inside the receiver by strong signals, and they scale faster than the signals that produce them: a third-order product changes three decibels for every decibel of input. If padding the input by 10 dB drops a feature by 30 dB, the feature was never in the air. This is also the practical reason the input intercept figures of section 10 matter: they are the specification that tells you how much signal a receiver tolerates before it starts inventing.

50Selection, by the job rather than by the datasheet

The method below is deliberately vendor-neutral. It produces a requirement that several suppliers can answer honestly, which is the only kind worth writing.

1. Start with the shortest event you must not miss. That fixes whether you need real-time capture at all, and if you do, equation (8) fixes the largest transform you may use.
2. Then the widest span you must hold at once. Instantaneous bandwidth, not tuning range. This is the number most often confused with the next one.
3. Then the highest frequency. Above roughly 6 GHz the integrated-transceiver class thins out and a different architecture is required, so this question can eliminate a whole tier of the market in one line.
4. Then ask whether the answer has to be defensible. If somebody outside your team must accept the number, you need a stated amplitude bound and a calibration date, and you have just moved into the fourth class of device.
5. Then transmit, duplex and coherence, each stated as a requirement about the relationship between channels rather than a count of them.
6. Then timing. Reference accuracy, holdover and timestamp accuracy, asked separately.
7. Only then the software. Which framework, which abstraction layer, and whether the processing investment survives a change of radio.

Figure 18 is the method as a picture; Table 17 turns it into a form you can put on a requisition.

51Ten jobs, and what decides each one

Selection is easier when the job is named. Each entry below is a problem this class of instrument is bought for, the property that actually decides it, and the Berkeley Nucleonics application note where the arithmetic is derived rather than asserted.

Table 18 is that map.

The ten, in a sentence each

The table is the index. These are the arguments, one paragraph each, for a reader deciding which note to open.

• Programming an analyzer as a platform. The decision gets argued on noise figure and is actually about whether the samples are reachable and whether the calibration reaches them too. An uncorrected board hands a flowgraph a fraction of full scale; a calibrated instrument hands it a level in dBm with a stated bound.
• Wide-area monitoring with automated classification. The receiver has become a sensor feeding a statistical estimator, and the estimator inherits every property of the data. A swept receiver at 1 MHz resolution across 40 GHz with a 100 ms revisit dwells 2.5 microseconds per cell, which is one part in 40,000 of the revisit, so that is the fraction of any emitter's transmissions arriving while it is looking, whatever their rate. No model recovers what that discarded.
• Detecting unmanned aircraft and their operators. The control link is present in any one channel for milliseconds, and the operator is found by direction-finding the uplink rather than the aircraft. What a visiting receiver reports as empty is a function of when it looked.
• Commissioning a satellite earth station. The deliverable is the file the visit produces. An interference claim arrives months later as an assertion, and the commissioning archive is the only timestamped, instrument-traceable evidence of a compliant state.
• RF record and playback. A replay can never be cleaner than its capture: the analyzer's noise, spurs and imbalance are burned into the file and are indistinguishable from signal to the generator. The capture instrument sets the ceiling.
• UAV-borne antenna measurement. Far field begins near two aperture-squared over wavelength, and at exactly that distance the phase error across the aperture is one sixteenth of a wavelength. A pattern point is a statistic over a population of pulses, not a single reading.
• EMF safety and RF exposure. Limits are written in field strength or power density and an analyzer reports power at a connector, so the job is a conversion with an uncertainty budget behind it, of exactly the kind section 46 works through.
• 5G, Wi-Fi and Bluetooth coexistence. Coexistence problems are between-standard problems, so an instrument that measures one standard at a time cannot see them. Holding 100 MHz gap-free puts the interferer and the victim in one acquisition against one amplitude reference.
• EMC pre-compliance on your own bench. No ICX specification table publishes an EMC measurement function, so this is ordinary spectrum work plus arithmetic you do yourself, not a compliance mode. The arithmetic that decides whether a scan is honest is dwell time against the intercept relation. Running a confirmation scan at search speed is the commonest way a pre-compliance result turns out optimistic.
• Designing a receiver into your own product. In a distributed radio-location node the receiver's own timing contributes a tiny fraction of the position-error variance while time transfer, inter-channel matching, enclosure temperature and site geometry contribute nearly all of it. One nanosecond is 0.2998 m, and geometry multiplies whatever survives.

52Troubleshooting, by symptom

Table 19 is arranged by symptom, because a symptom is what you have when you open this page. The third column is the point of it: most of these symptoms have several plausible causes, and the discriminating check is what stops an afternoon going on the wrong one.

53Buy, build, or integrate

One decision sits above the selection method, and getting it wrong is more expensive than any subsequent choice. Three routes exist to a receiving system, and the arithmetic that separates them is not about radio performance.

Table 20 sets the three routes side by side.

The crossover is a volume, and computing it beats arguing about it. A fixed development and qualification cost spread over a lifetime volume meets a flat unit price at one quantity, and BNC-AN-110 derives that crossover in full for an embedded monitoring product. The result most teams find surprising is how high the crossover sits once calibration and obsolescence engineering are counted honestly.

54Compliance, export and supply chain

The regulatory work in section 53 is the item most often left out of a schedule, so it is worth naming what it consists of. Five separate regimes apply to a receiving product, they are administered by different bodies, and satisfying one says nothing about the others.

Table 21 separates them.

The evaluation-kit exclusion deserves the emphasis. A development board is exempt because it is a development board. Putting it in an enclosure and selling it makes it a product, and the product is assessed on its own, from scratch, by whoever puts their name on it. That work does not transfer from the board vendor and it is not small: it is a test campaign, a technical file, a declaration of conformity, and a person willing to sign it.

The Cyber Resilience Act is the newest of the five and the one most likely to be missed, because it does not look like radio regulation. It reaches software, and a receiver is a software product with an antenna on it. Two features make it consequential for a buyer. The reporting duty applies to products already on the market, so it is not something a supplier can defer until the next design. And it runs for the whole supported life of the product, which turns a question most datasheets never answer, how long this thing will be supported, into a regulated commitment. Ask for the support window in writing, and ask who issues firmware updates and how they are signed.

55Provenance, and the questions procurement will ask

Where a receiver was designed and built has become a purchasing criterion in its own right, independent of performance. Under Section 889 of the FY2019 National Defense Authorization Act, federal agencies may not procure covered telecommunications equipment as a substantial or essential component of any system. Part A, the procurement prohibition, took effect on August 13, 2019; Part B, which bars agencies from contracting with an entity that uses such equipment, took effect on August 13, 2020. Part B is the one that surprises people, because it reaches past what you are selling to what you are running.

The practical consequence for a receiver is that the answer has to hold at component level, not at brand level. A platform assembled from modules is only as clean as its least documented module, and a bill of materials nobody has read is not an answer. Three questions establish more than any assurance does:

• Who designed the radio, and where is it manufactured? Not where the box is assembled, which is a different question and often a different country.
• Can you produce a bill of materials to component level, and will you flag changes? A supplier who cannot has not been asked before.
• What is the support and repair path, and who performs the calibration? An instrument returned across a border for recalibration is a supply-chain fact, not a service detail.

So that we are held to the same standard: Berkeley Nucleonics is a California company, and the ICX line is supported and calibrated through it. That answers the third question and part of the first. It is deliberately not an answer to component-level origin, which is a bill-of-materials question and belongs in writing against a specific configuration rather than in a handbook. Ask us for it the same way you would ask anyone, and treat a brochure paragraph from any supplier, including this one, as the start of the conversation rather than the end of it. Nothing here is a claim about anybody else's provenance.

56Reading somebody else's datasheet honestly

This handbook cannot compare products for you, and would not be worth much if it did, because the comparison you need is against your requirement rather than against a league table. What it can do is give you the questions that make any datasheet answer for itself.

• Find what condition each number carries, and treat an unconditioned number as absent. A noise figure without a frequency, a sweep speed without a resolution bandwidth, a phase noise without an offset: each is a number that cannot be reproduced and therefore cannot be checked. This applies to our datasheets exactly as much as to anyone's.
• Look for the pairs that should both be present. Sample rate and usable bandwidth. Burst and sustained. Resolution and effective bits. Preamp-on and preamp-off damage limits. Where only the flattering half of a pair appears, that is a choice somebody made.
• Check whether a range is a specification or a coverage claim. Many devices publish a frequency range far wider than the range over which any performance figure is stated. The question to ask is not how wide the range is but over what part of it the accuracy holds.
• Count the tiers with specification tables. A family advertised in five frequency tiers with published tables for two has told you something. This is worth applying to the ICX-FieldHawk too: only the 9.5 GHz and 40 GHz tiers carry published tables, and nothing in this handbook is written against any other.
• Ask what is specified about relationships, not just about parts. Between channels, between units, between the calibration and the samples. Section 18 lists these and they are consistently the least published and the most consequential.

57Measurement practice that applies whatever you bought

1. Let it warm up. Published accuracy figures assume it. Measurements from a cold start are covered by no specification on any datasheet.
2. Set the reference level deliberately. Amplitude accuracy, intermodulation performance and displayed noise floor all move with it, and an automatic reference level during a comparison invalidates the comparison.
3. Read back every setting you request. A sample rate, a bandwidth or a frequency is a request, not an assignment: hardware has step sizes, and a filter designed against the value you asked for and fed the value you got produces a slow error nothing downstream catches.
4. Record the state, not just the trace. Center frequency, span, sample rate as read back, resolution bandwidth and video bandwidth, decimation, transform size, window, detector, trace mode, reference level, attenuation, preamplifier state, antenna and its factor, cable assembly, and a timestamp. A capture without that sidecar is a pile of integers.
5. Fix the cable and leave it fixed. Loss in flexible assemblies changes when the cable moves, so swapping one mid-campaign is an uncalibrated step in the middle of the record.
6. Record the resolution bandwidth with every noise figure. A displayed average noise level per hertz becomes a real floor only after ten times the log of the bandwidth is added, and the two get confused constantly.

58Two worked selections

The method is easier to trust when it has been run. These are two real shapes of requirement, worked through the questions of section 50 to a class of device rather than to a product.

A university teaching and research laboratory

The job is teaching modulation and protocol work to students, plus occasional research captures below 6 GHz that end up in papers. Budget is per-seat rather than per-instrument.

• Shortest guaranteed event: not a constraint. The signals are generated in the lab and repeat on demand, so intercept probability does not decide anything.
• Widest span held at once: 20 MHz covers every waveform on the syllabus.
• Highest frequency: 6 GHz. Everything taught lives below it.
• Does the answer have to be defensible? For teaching, no. For the research captures that reach a paper, yes, and this is the split that matters.
• Conclusion: development-class transceivers for the teaching seats, where the open ecosystem and the price are exactly right, and one instrument-grade receiver shared across the group for anything that gets published. Buying instruments for every seat would waste money on a column nobody grades; buying none would mean every published amplitude rests on an uncalibrated board.

A distributed monitoring network across a site

The job is unattended nodes across a large industrial site, reporting occupancy and flagging unexpected emitters, expected to run for years and to be extended later.

• Shortest guaranteed event: short. The emitters of interest are intermittent, which makes intercept probability the first specification rather than an afterthought, and equation (8) then fixes the largest transform each node may use.
• Widest span held at once: as wide as the site's allocation, held continuously rather than swept, for the reason Figure 8 shows.
• Highest frequency: set by the site, and if it reaches past 6 GHz that decision has already eliminated a tier of the market.
• Defensible? Yes. An occupancy claim that leads to an enforcement conversation has to survive being questioned.
• Timing: the deciding requirement, and the one most likely to be under-specified. Ask for holdover separately, because the nodes will lose their discipline.
• Transport: name a format. This system is explicitly expected to outlive its hardware, which is the case section 23 was written for.
• Conclusion: instrument-grade modules with a published intercept relation, a stated amplitude bound, a timing option and a documented interface. And a written answer on transport format before committing, since that is the requirement most likely to be unmet, including by us.

59What lives where: a working band reference

Half of the work in an unfamiliar capture is knowing what you are probably looking at before you identify it. The allocations below are the ones that account for most of what a wideband receiver encounters below 8 GHz, and carrying them in your head beats looking them up each time.

Table 22 is the short version.

Above 8 GHz the naming changes from allocations to the IEEE letter bands, and those are worth knowing because radar, satellite and point-to-point work is described in them almost exclusively.

Table 23 gives the letters.

60Where the bands are going

A reference that only describes today ages badly, and a receiver bought today will be in service for years. Two movements are worth writing into a specification now.

The upper mid-band. 3GPP calls 7.125 to 24.25 GHz FR3, and it is the most active piece of spectrum policy in the world at the moment. Resolution 256 (WRC-23) set studies running on 4400 to 4800 MHz, 7125 to 8400 MHz or parts of it, and 14.8 to 15.35 GHz, and those studies feed WRC-27 agenda item 1.7, which is where identification for mobile use would actually be decided. WRC-23 scheduled the question; it did not answer it. It is the leading candidate spectrum for 6G because it sits between the coverage of the sub-6 bands and the capacity of millimeter wave.

Sharing rather than allocation. The second movement is structural. CBRS put three tiers of user in one band under an automated coordinator, and 6 GHz put unlicensed devices under automated frequency coordination in the presence of incumbent links. Both replace a static allocation with a database that changes who may transmit, where, and when. For a receiver this changes the job: occupancy becomes a function of time and place by design, not by accident, so a single visit to a band no longer characterizes it even in principle, and evidence about who was transmitting has to carry a timestamp and a position to mean anything at all.

61Two more movements a receiver will meet

Signals in the cellular bands may now be coming from orbit. 3GPP Release 17 brought non-terrestrial networks into the same air interface as terrestrial cellular, and in 2024 the United States adopted a supplemental-coverage-from-space framework under which satellite operators transmit in a mobile operator's own terrestrial spectrum, as a secondary user, to reach ordinary handsets. Constellations are doing this today in PCS and AWS spectrum.

And the fronthaul argument of section 23 has a second front. Open radio access network work disaggregates a base station into a radio unit and processing units from potentially different vendors, connected by an open fronthaul interface that carries digitized radio across an Ethernet network. It is the same structural story this handbook told about digital IF transport, in a much larger market: once an analog interface becomes packets, interoperability stops being automatic and has to be specified, tested and certified. The practical consequence for test equipment is that the thing under test is increasingly a packet stream rather than a cable, and an instrument that can only measure at an antenna port measures one half of a modern radio site.

62Quick reference

Table 24 is the page to photocopy.

63A specification you can send to several vendors

The method in section 50 produces requirements. This section turns them into wording, because a requirement phrased badly gets answered by everyone and distinguishes nobody. Each line below is written so that a supplier must either give a number or say it does not publish one.

1  ACQUISITION
1.1  Instantaneous bandwidth digitized at one instant:        ____ MHz
1.2  Usable bandwidth within that span, stated separately:    ____ MHz
1.3  Guaranteed intercept at 100% probability, with the
     transform size it assumes:                               ____ us at N = ____
1.4  Sustained record rate to host, distinct from burst:      ____ MHz sustained
                                                              ____ MHz burst

2  AMPLITUDE
2.1  Absolute amplitude accuracy, as a bound, with the
     frequency range and conditions it holds over:            +/- ____ dB
2.2  Displayed average noise level, with frequency and RBW:   ____ dBm/Hz
2.3  Maximum input, preamp off and preamp on:                 ____ / ____ dBm
2.4  Input intercept, third order, with reference level:      ____ dBm

3  TIMING
3.1  Reference accuracy, disciplined:                         ____ ppm
3.2  Holdover over a stated interval after discipline is
     removed:                                                 ____ ppm over ____
3.3  Timestamp accuracy against 1PPS:                         ____ ns

4  MULTI-CHANNEL  (omit if single channel)
4.1  Specified relative phase or delay between channels:      ____
4.2  Common reference and hardware trigger provided:          yes / no

5  INTERFACE
5.1  Documented programming interface, languages named:       ____
5.2  Transport format, and whether VITA 49 or DIFI:           ____
5.3  Driver for an open hardware abstraction layer:           yes / no
5.4  Is device calibration applied before samples cross
     the interface:                                           yes / no

Two notes on using it. Send the same form to everyone, including us, and compare the shape of the blanks rather than the numbers: a supplier who leaves four lines empty has told you which class of device you are being offered. And do not weight the lines before the answers come back, because the temptation afterwards is to weight them toward whoever answered best.

64Standards worth knowing by number

Table 25 collects the standards this handbook refers to, with what each one governs. A specification that cites a standard by number is answerable; one that describes a capability in prose is not.

65A reproducibility checklist

Whatever class of device you bought, a capture that cannot be reproduced is a capture that will be argued about. Copy this into a log and fill it in. A capture without these entries is not reproducible, whatever else is true of it.

1. Instrument model and serial number, and its calibration date if it has one.
2. Ambient temperature, and warm-up elapsed before the first reading.
3. Center frequency, span, resolution bandwidth and video bandwidth.
4. Sample rate as read back from the device, not as requested.
5. Decimation, transform size, window, detector and trace mode.
6. Reference level, attenuation and preamplifier state.
7. Whether device automatic gain control was on or off, stated explicitly.
8. Antenna, its antenna factor at the frequency of interest, and the cable assembly used.
9. Software versions: framework, abstraction layer, and every out-of-tree module.
10. Overflow or dropped-sample events with timestamps, or an explicit statement that none occurred.
11. The result of an amplitude check against a known level, with the date it was taken.

66Definitions, symbols and further reading

Aliasing

The folding of signals outside the digitized span into it. Once folded they are indistinguishable from signals that were really there, which is what the anti-alias filter exists to prevent.

Antenna factor

The conversion, in dB per meter, between the field at the antenna and the voltage at its connector. Required before any receiver reading becomes a field strength.

Back pressure

The propagation of a stall upstream through a processing graph when a block cannot consume as fast as its input arrives.

Coverage factor

The multiplier applied to a standard uncertainty to give a stated confidence. A factor of two is conventionally quoted as about 95 percent, exactly so only for a near-normal combined distribution.

Detector

The rule that reduces the many samples behind one displayed point to a single value: peak, sample, average or RMS. Only an RMS detector, or an average detector that averages power rather than decibels, reports correct power for a noise-like signal; log averaging under-reads noise by 2.51 dB.

Digital down-conversion

Shifting a band to zero frequency by multiplying the samples against a numerically generated oscillator, filtering, and only then decimating. The arithmetic equivalent of a mixer, and what a published decimation range specifies.

Dilution of precision

How much the geometry of a set of sensors or satellites multiplies a timing error into a position error. A value of three turns 10 m of range uncertainty into about 30 m of position uncertainty.

Direct sampling

Digitizing the antenna signal without analog down-conversion. Selectivity becomes arithmetic, and the converter sets the frequency ceiling.

Effective number of bits

Converter resolution derived from measured signal to noise and distortion, rather than from the nominal bit count. Always lower.

Equivalent noise bandwidth

The width of the ideal rectangular filter that would pass the same noise power as the window in use. It converts bin spacing into resolution bandwidth, and it is 1.9761 for the Blackman-Nuttall window and about 3.77 for a flat-top.

Error vector magnitude

The distance between where a received symbol landed and where it should have landed, as a fraction of the constellation. It measures modulation quality and is set by noise, phase noise and loop bandwidths rather than by calibration.

Gap-free

Every input sample reaches a transform, with no dead time for retune or processing. A property of the acquisition, not of the display.

Holdover

How far a reference drifts once external discipline is removed. Governed by aging and temperature coefficient, and usually unpublished.

Instantaneous bandwidth

The span digitized at one instant. Distinct from tuning range, which is covered by moving the tuner.

Intercept, guaranteed

The observation length over which an event is captured at full amplitude with certainty. A property of a real-time engine, and unrelated to the distortion sense of the word below.

Intercept, third-order

The extrapolated input level at which a receiver's third-order distortion products would equal the signals producing them. A measure of how much signal the receiver tolerates before it invents one.

Intermodulation product

A spurious signal generated inside a receiver by two or more strong inputs. A third-order product changes three decibels for every decibel of input, which is the test that identifies it.

Minimum detectable signal

The noise floor referred to the input: thermal noise density at 290 K, plus the noise figure, plus ten times the log of the measurement bandwidth.

Preselection

Filtering ahead of the receiver so that emitters outside the band of interest do not consume its dynamic range.

Probability of intercept

See Intercept, guaranteed. The two names are used for the same quantity, and the guarantee form is the one a specification can be written against.

Quadrature imbalance

A gain or phase mismatch between the in-phase and quadrature paths of a zero-IF receiver, seen as an image at the negative of a signal's offset.

Rectangular distribution

The assumption applied to a bound quoted without a distribution, contributing the square of its half-width over three to a variance budget.

Resolution bandwidth

The width of the filter or transform bin each displayed point represents. It sets the noise floor, so an amplitude quoted without it is not a measurement.

Spur reject

An instrument mode that suppresses internally generated spurious responses, at a cost in sweep speed. It is a standing condition on every published ICX figure quoted here, so the same instrument with it bypassed is quoting a different number.

Spurious-free dynamic range

The range between a signal and the largest spurious response the receiver generates from it. Distinct from noise-limited dynamic range.

TCXO and OCXO

Temperature-compensated and oven-controlled crystal oscillators. The oven costs power and warm-up time and buys roughly an order of magnitude in stability and considerably more in holdover, which is why it is an option rather than the standard fit.

Traceability

The unbroken chain from a reading to a national or international standard, with a stated uncertainty at each step. What separates a measurement from a reading.

U-shaped distribution

The distribution of a mismatch error, whose phase is uniform so its amplitude concentrates at the extremes. It contributes the square of its half-width over two, half again more than a rectangular assumption would.

Zero IF

Mixing the wanted band directly to baseband as I and Q. The architecture behind most integrated transceivers, and the source of three characteristic impairments.

Sources cited in this handbook

• Analog Devices, AD9361 RF Agile Transceiver product data. Receive local oscillator 70 MHz to 6.0 GHz, transmit 47 MHz to 6.0 GHz, 2 by 2 channels. analog.com/en/products/ad9361.html
• Analog Devices, AD9364 product data. The 1 by 1 variant, sharing a register map and package with the AD9361. analog.com/en/products/ad9364.html
• DIFI Consortium, formed under the IEEE Industry Standards and Technology Organization to advance interoperability between space and ground systems and to prevent vendor lock-in. dificonsortium.org
• IEEE-ISTO Std 4900-2021, Digital IF Interoperability Standard: a constrained profile of VITA 49.2 that defines what is required rather than what is possible. Tutorial material at github.com/DIFI-Consortium/DIFI-Certification
• ANSI/VITA 49, the VITA Radio Transport standard for carrying digitized RF with its metadata.
• GNU Radio project documentation, for the framework and hardware ecosystem described in Part III. wiki.gnuradio.org
• SoapySDR, the vendor-neutral hardware abstraction layer named in Part III. github.com/pothosware/SoapySDR
• Signal Metadata Format (SigMF), the sample-file plus JSON-sidecar convention recommended in section 40. github.com/sigmf/SigMF
• AMD Zynq UltraScale+ RFSoC and Versal RF series product data, for the direct-sampling converter rates, the 7.125 GHz input range and the December 2024 announcement quoted in section 6. amd.com
• 3GPP Release 17 for non-terrestrial networks, and the FCC supplemental coverage from space framework of 2024, for section 61. 3gpp.org and fcc.gov
• O-RAN Alliance open fronthaul specifications, for the disaggregated radio access network discussed in section 61. o-ran.org
• Federal Communications Commission, Report and Order of 22 July 2026 on the upper C-band, for the 3.98 to 4.14 GHz reallocation in section 59. fcc.gov
• Resolution 256 (WRC-23) and WRC-27 agenda item 1.7, for the upper mid-band study bands in section 60. itu.int
• Directive 2014/53/EU (radio equipment), Directive 2014/30/EU (electromagnetic compatibility) and Regulation (EU) 2024/2847 (the Cyber Resilience Act), for the compliance table. eur-lex.europa.eu
• Section 889, FY2019 National Defense Authorization Act, and the implementing Federal Acquisition Regulation rules, for the provenance section. acquisition.gov
• Berkeley Nucleonics, ICX-FieldHawk handheld, rugged and USB datasheets, for every ICX figure in this handbook.

Companion documents

• BNC-HB-001, SDRs in the real-time spectrum analyzer field. The deep treatment of SoapySDR, GNU Radio flowgraphs and the calibration boundary that Part III of this handbook summarizes.
• BNC-AN-102, wide-area spectrum monitoring, for the intercept arithmetic behind section 32.
• BNC-AN-105, RF record and playback, for the two-instrument split named in section 45.
• BNC-AN-110, designing a receiver into your own product, for the timing and geometry arithmetic behind Part II.

67What to take away

Six things, if the rest is forgotten.

• The market clusters because it shares silicon. One integrated transceiver family covers 70 MHz to 6.0 GHz and underlies a large share of the products on sale, which is why so many unrelated devices stop at the same frequency and carry the same three impairments. A specification shared by a dozen products is not a differentiator.
• Sample rate is not bandwidth, and neither is tuning range. Four different numbers get quoted as a span, and only the narrowest of them governs a measurement.
• The bit count is a ceiling, not a performance figure. Effective bits after noise and distortion are always fewer, and the gap between a consumer converter and an instrument one is the largest technical difference in the landscape and the least visible on a datasheet.
• A reading is not a measurement. The dividing line in this market is whether an absolute amplitude bound exists, and it is a property of hardware and its traceability chain that no software layer above can supply.
• A guarantee you can evaluate beats a claim you cannot. A published intercept relation lets a buyer price a configuration the vendor never considered; a marketing phrase about real-time performance does not.
• Ask about relationships, not just parts. Between channels, between units, between the calibration and the samples. These are consistently the least published fields and the most consequential ones.