| Episodes | 64 |
| Mentions | 79 |
| Cited here | 38 |
| First — last | #16 — #727 |
| Top guests | Greg Charvat, Ben Krasnow, Jonathan Beri |
| Related | analog to digital converter · datasheet · software defined radio · usb · bluetooth |
Bandwidth is the width of the range of frequencies that a signal occupies, or that a system can pass, transmit, or capture.[463] The term is literal — it denotes the width a signal occupies on the band — and it functions across electronics as a unifying constraint: in instruments it sets what can be measured, in radio systems it sets what can be received or resolved, and in digital links it bounds what can be transferred.[463][214] Because bandwidth trades against nearly every other specification — noise, power, distortion, resolution, efficiency and cost — it is one of the principal currencies of engineering design, spent in one place to buy performance in another.[65][407] It is also a finite resource at the level of the electromagnetic spectrum itself, which shapes the economics of wireless networks and the lifespan of devices that depend on them.[462]
Concept and intuition
The word is most often first encountered as an abstraction, and seeing a signal’s occupied width directly on a spectrum waterfall display is what converts the term from vocabulary into concept.[463] Interactively narrowing and widening a receiver’s bandwidth while watching the selection window move over the spectrum teaches the idea faster than the underlying mathematics, which is an argument for placing inexpensive software-defined radios in front of anyone learning signals.[381]
Test and measurement
Oscilloscopes and sampling
Nyquist’s factor of two between sample rate and bandwidth is the mathematical floor, not the engineering figure. An oscilloscope with a Gaussian frequency response requires a sample rate of about 2.3 times its bandwidth to be guaranteed free of aliasing — 230 megasamples per second for a 100 MHz front end — while a factor of ten is the usual rule of thumb for reproducing a waveform that looks correct.[570] An instrument quoting 100 megasamples per second against 30 MHz of bandwidth is running only about three times oversampled: enough that it will not report a wrong frequency through aliasing, but not enough to reproduce the shape of the signal faithfully.[339]
Two developments made digital oscilloscopes practical: real-time sampling, under which ten times the sample rate bought ten times the bandwidth, and subsequently deep sample memory, the second of these being the one that changed how the instrument is used day to day.[72] In USB oscilloscopes, resolution has never been the scarce quantity — 12- and 16-bit converters are commonplace — and bandwidth is what the purchase price actually buys.[677] When comparing acquisition hardware, the bandwidth figure is the one that matters and the sample count follows from it; additional resolution bits are real only to the extent that the noise floor permits their use.[193] Because an entire 800 MHz oscilloscope front end now fits in a single chip, instruments across a vendor’s range frequently share identical hardware and differ only in what the firmware permits.[654]
Probes and front ends
Passive oscilloscope probes run out somewhere around 300 MHz, arguably 500 MHz, so beyond that point the probe rather than the oscilloscope sets the usable bandwidth of the measurement; buying more oscilloscope without changing the probe buys nothing.[347]
In precision instruments, the front-end topology itself can cap bandwidth. The conventional cascaded resistive divider in a multimeter input cannot be held flat to high frequency, because each junction in the chain adds a node with its own stray capacitance; it is the topology, not the component values, that imposes the limit.[180] In Dave Taylor’s precision meter work, where the board material itself contributed the limiting capacitance, the remedy was to remove it — cutting the FR4 laminate away around a high-impedance divider to take the dielectric out of the circuit entirely — and bandwidth drove more prototype-board iterations than any other requirement in the design.[180] Meeting a headline bandwidth specification is in general a larger commitment than the specification line suggests, because serious analog work requires reading the performance curves across the full band of interest rather than relying on the single figure at the top of the datasheet.[168]
Signal integrity
Clock frequency does not bound the spectrum a signal occupies; rise time does. A 32 MHz clock with a 100-picosecond edge places spectral content out to roughly 3 GHz, which lands squarely on any nearby receiver listening for signals at the −100 dB level.[252] On mixed-signal boards the corresponding design rule is to use the lowest clock frequency and the longest rise time the design tolerates: a handful of ordinary layout rules suffices to around 100 MHz, but past thirty-odd megahertz with a radio on the same board, crosstalk and shielding become the central design problem.[252] Propagation speed also enters the timing budget: a signal travels down a trace at the speed of light in the laminate rather than in air, and the dielectric constant of about four for FR4 slows propagation by its square root, which is what makes trace length a timing quantity rather than a mechanical one.[252]
Analog circuit design
Bandwidth functions in component selection as a currency that buys other specifications. A converter limited to 4 kHz — too narrow even for audio — can offer distortion performance that nothing else matches, together with extremely low power consumption for the bandwidth it does deliver.[65] Conversely, a newer part advertised as higher bandwidth can be worse where it matters: increased bandwidth paired with a reduced slew rate is a real combination, and only reading the entire datasheet rather than the headline figure reveals it.[727] Substituting a faster transistor into a working design can likewise degrade it, because the higher-bandwidth device amplifies noise that the slower part had quietly ignored; a higher-bandwidth part is therefore not automatically a drop-in improvement.[290]
At the edge of what a single operational amplifier can do, a composite amplifier places a second device inside the feedback loop of the first so that their strengths combine — for example a very low-noise, low-offset input stage with a high-drive output stage able to swing a coaxial cable. Cascading two amplifiers in series is a different arrangement that buys bandwidth or drive but not this mix-and-match of properties.[660]
In control-loop engineering the word takes on a related but distinct meaning: loop bandwidth describes how fast the loop can respond. A right-half-plane zero, characteristic of boost-derived converter topologies, limits how fast the loop can be made and can therefore rule out topologies that otherwise look attractive on cost.[566]
Radio and communications
Receiver architectures
A superheterodyne receiver discards half of the available information at down-conversion, whereas an in-phase and quadrature (I/Q) direct-conversion receiver keeps both components, yielding twice the usable bandwidth and enough information to perform image rejection digitally rather than with aggressive analog filters.[52] Filtering itself can be done in either the time or frequency domain, and the crossover point arrives when a time-domain filter needs so many taps that building a good one becomes difficult, at which point moving into the frequency domain is the cheaper path.[52]
Spread spectrum and software-defined radio
The two common spread-spectrum families place the implementation burden in different places: frequency hopping moves the signal around and requires a receiver that retunes to follow it, while direct sequence stays on one frequency but occupies a wide band, so the receiver must capture a wide slice at once.[352] Instantaneous visible bandwidth on such platforms is limited — under 20 MHz on one popular design — so covering a six-gigahertz tuning range requires sweeping, and moving the retuning into device firmware removes the USB round-trip latency that would otherwise dominate every step of the sweep.[352]
A related capture-bandwidth argument applies to side-channel analysis: attacking a gigahertz-clocked device requires only twenty or thirty megahertz of capture bandwidth, because a down-converter moves the band of interest to where an ordinary digitiser can reach it.[239]
Data rate, power and coverage
Moving to millimetre-wave carriers does not by itself deliver more throughput, because the converters rather than the radio set the limit: an enormous carrier frequency is of no help if the system can digitise only a 50 MHz slice of it.[483] High data rates also carry a power cost that does not shrink with the radio — a gigabit per second into a handset implies a baseband processor clocking at gigahertz, and the resulting dissipation is incompatible with a device held in the hand.[483] The high-bandwidth portion of 5G is millimetre wave and trades range for capacity, which is why it implies transceivers distributed on poles throughout a city rather than a few tall masts; most of what is deployed under the 5G label is additional spectrum in the existing lower bands.[569]
Cellular categories intended for low-power devices are defined by narrowing the occupied band: each step down the ladder squeezes the signal into less spectrum for less throughput, and the narrowband tier goes further still with different modulation and less data again.[509] The converse also holds. Sending very little data buys receiver sensitivity, because a low rate leaves time for processing gain. In Larry Sears’s utility-metering work, a meter needed to convey roughly one bit per hour, which permitted simple FM modulation, a cheap receiver, and a transmit power low enough to meet emissions limits without effort.[109] At the other extreme of packing density, Trammell Hudson’s digital amateur-radio mode occupies about 15 Hz of bandwidth, fitting many simultaneous users onto one voice channel; shaping every transition with a cosine-squared ramp is what keeps each user from splattering into the others.[463]
Bandwidth as resolution and robustness
In a chirped radar, the swept bandwidth is the range resolution, not the carrier frequency: a well-known amateur coffee-can radar design sweeps 80 MHz of bandwidth, and that figure is what sets its range resolution.[214] Similarly, a resonant transducer is narrowband by nature, and bandwidth is bought by damping it — a direct trade of efficiency for frequency span.[407] Frequency-modulated recording illustrates a third use of the same lever: placing the carrier only a little above the baseband it carries — roughly 5 to 10 MHz of carrier for 4 MHz of video — makes the recovered signal highly tolerant of tape dropouts and level variation.[133]
Discrimination can also be built from bandwidth directly. A sound-activated switch discriminates on four properties at once — frequency through a bandpass filter, bandwidth, duration and amplitude — and loosening any of them is what turned cheaper versions of such products into units that triggered on their own.[690]
Spectrum economics
Spectrum is fixed, so a carrier’s incentive is always to reclaim an old allocation and repurpose it for a denser standard; this makes any long-lived industrial device dependent on a legacy network a dated asset regardless of how well the device itself works.[462] For sensor traffic, the oldest cellular generations were always adequate on bandwidth and inadequate only in device-side power consumption, and the constraint that actually strands such designs is that the networks are being switched off to reclaim the spectrum.[678] In satellite service, throughput is set by contention with nearby users rather than by remoteness, so an isolated location does not receive a private link; measured rates of 400 to 800 kilobits per second are common enough to render modern web pages barely usable.[713]
Digital and embedded systems
Aggregate data rate, rather than processing power, is frequently the binding constraint in embedded systems. A single processor driving many shift-register display modules fails on the aggregate data rate; giving each module its own small controller and letting the controllers exchange data is what makes the system scale.[16] For a single display, driving it over a parallel interface with direct memory access yields far more bandwidth than the serial interface most projects reach for by default, at no cost in processor time once the transfer is set up.[356] Where computation must occur before the host can cope with the data rate, programmable logic is the appropriate tool, and the discipline of leaving it out when the data rate does not require it is a mark of good design rather than a limitation.[214]
Constrained networks
Protocol choice dominates bandwidth consumption on constrained links. Swapping a web-oriented request protocol for one designed for constrained devices saves twenty to forty percent of both bandwidth and battery for exactly the same message, because the framing omits overhead a small device never needed.[526] A team experienced in scalable cloud services can burn a fleet’s entire monthly cellular data allowance in a single day, because nothing in that background prepares designers for a link where each message carries a cost; the protocol choice is effectively a hardware decision, not a server one.[526] At the consumer end, a domestic appliance consuming a gigabyte per day is a design smell rather than a feature, and the usual explanation is verbose human-readable logging left enabled in shipping firmware.[541] Bit-level frugality is a response to a constraint rather than a virtue in itself: a deep-space probe packs every bit and pays for it in decoding complexity at the receiving end, while a device on a cheap link can reasonably send plain text — the error is applying either habit where the other belongs.[541]
Matching the reporting rate to the physical rate of change of the measured quantity is usually the largest single saving available in a battery-powered sensor, since quantities such as temperature change slowly while customers ask for continuous high-rate telemetry.[153] The same logic applies at larger scale: downlinking imagery that turns out to be cloud cover wastes bandwidth, operational cost and energy in equal measure, which is the clearest case for deciding on board what is worth transmitting rather than filtering after the fact.[517]
A further subtlety is that retransmissions present as a halving of throughput rather than as an error, which makes a marginal link look like a merely slow one and sends fault investigation in the wrong direction.[554]