FASTSTREAM Talk to us
PlatformsFaststream SiliconFaststream RadioFaststream VisionConnected EdgeFaststream SecureMobility & Rail
ProductsSemiconductor IPWireless & RANEdge & GatewaysTracking & IdentificationSoftware & FrameworksConnected Systems
Technology5G protocol stackWireless and RF architectureBaseband and low PHYForward error correctionControl and data planeHigh-speed interfacesFirmware and bootSilicon root of trustSoftware-defined vehicleAutomotive OTAFunctional safety
AIAI Engineering ServicesEdge AI & Embedded MLComputer Vision EngineeringSensor Fusion & PerceptionAI Silicon & AccelerationMLOps for DevicesAI Visual InspectionPredictive MaintenanceDriver MonitoringVideo Analytics & Safety
SolutionsSemiconductorIndustrial AIConnected ProductsAsset TrackingBluetooth AoA RTLSWearable TrackingAutomotive & MobilitySmart InfrastructureSecure IdentityWireless & SatellitePrivate 5GSmart WashroomsFuel ManagementSmart BuildingsWorker SafetyEnergy MonitoringSmart AgricultureSmart CityAutonomous PlatformsAssembly AutomationLiDAR Rail SafetyHardware Wallet
IndustriesSemiconductorTelecommunicationsIndustrial & ManufacturingAutomotive & MobilityTransportation & RailAerospace & DefenceHealthcare & MedicalEnergy & UtilitiesOil & GasRetailConsumer ElectronicsMedia & EntertainmentSmart Infrastructure & IoT
ServicesSystem Integration overviewASIC & SoC DesignRTL to GDSIIVerification methodologyDFT and silicon testLow-power designMixed-signal integrationDesign enablementFPGA DesignFPGA-to-ASIC ConversionAnalog, Mixed-Signal & RFHardware & High-Speed PCBEmbedded SoftwareCloud, OTA & Device ManagementManufacturing TransitionHow we engage
CompanyAbout FaststreamEngineering ExcellenceLeadership & OrganisationHow We EngageQuality & ComplianceStandards & EcosystemPartners & EcosystemTrust CentreLocations & DeliveryNewsroom & MediaCareersCase StudiesKnowledge CenterWhite PapersGlossaryNewsletterResources & Support
ContactStart a projectHow we engage
Talk to us
KNOWLEDGE CENTER

A knowledge center worth bookmarking.

Engineering writing rather than company news. Each piece answers a question an engineer actually asks, in enough detail to be useful, including the parts that argue against doing the thing.

ARTICLES

Seven pieces worth your time.

Latest
SILICONOctober 2026 · 7 min read

Why bugs still reach first silicon

Even great verification lets escapes through — the model is not the world, and simulation runs a blink of the cycles silicon does. Where they hide, and what the lab does.

Read →
ALL BEHAVIOUR OF THE PHYSICAL PARTEXERCISED BY SILICON OVER ITS LIFECOVERED BY THE VERIFICATION MODELSIMULATEDBEFORE TAPEOUTescapesWHERE THE ESCAPES LIVEAnalog & timing marginwhat the model idealised awayPVT corner behavioursilicon the simulator never sawReal interaction & sequencestraffic no testbench enumeratedTwo gaps compound: what the model left out,and what simulation had no time to reach.
TESTER PRESENTNO TESTER — THE CHIP IS ON ITS OWN, FOR YEARSshipWafer sortATE pattern setFinal testATE + BISTPower-on self-testevery key-onPeriodic self-testwhile runningLatent fault caughtbefore it mattersA factory test is a single event. The faults that hurt a safety-critical part are the ones that appear long after it — which only self-test is present to catch.
02SILICONOctober 2026 · 8 min read

Built-in self-test, and why it doesn't stop at the factory

A chip is tested once at the factory. Built-in self-test is how it keeps checking itself for the next fifteen years — and why safety makes that mandatory.

Read →
ONE WORKLOAD, FOUR ABSTRACTIONS — EARLIER IS FASTER, LATER IS MORE ACCURATEVirtual prototypeUntimed SystemC · earliest · fastestEmulationCycle-accurate · power & perfFPGA prototypeReal-time · real interfacesSiliconFinal validationSpeed & earliness decrease left to right; timing accuracy and cost increase. The discipline is keeping the same workload runnable across all four.
03SILICONSeptember 2026 · 9 min read

Hardware-software co-design in the AI era

When software defines the product, the chip becomes its execution platform. The barriers, the bottlenecks, and what actually helps.

Read →
WRONG IMAGING GIVES YOURIGHT IMAGING GIVES YOUSurface defectWashed out by glareLit at an angle, stands outA specular partA mirror of the light sourceDiffused, surface visibleFast-moving lineMotion-blurred, smearedStrobed, frozen sharpThree-dimensional flawInvisible top-downProfiled by structured lightLow-contrast markLost in the backgroundFiltered, made to popColour variationShifts with ambient lightReferenced, held stableThe model is only ever as good as the image. Most of the accuracy is designed into the light.
04VISIONAugust 2026 · 8 min read

No model recovers a defect the camera never caught

Machine-vision accuracy is won in the lighting, optics and exposure, not the model. Why a defect with no contrast cannot be detected, and why imaging is designed before the model.

Read →
TUNED HOT (HIGH SENSITIVITY)TUNED ON COST (CALIBRATED)Cosmetic fillet variationFlagged as a defectLearned as normalOperator workloadFloods of false calls to clearOnly real questions raisedTrust in the systemMuted by shift twoKept, because it is rightA genuine bridgeCaught, among the noiseCaught, and believedA rare critical defectLost in the false callsWeighted by consequenceThe decision thresholdSet to catch everythingSet on cost of each errorDetecting defects is easy. A low false-call rate at an acceptable escape rate is the engineering.
05VISIONAugust 2026 · 7 min read

The false call, not the missed defect, is what kills AOI

Automated optical inspection is judged on trust. Why flagging every cosmetic deviation loses the operator, and how fully learning a good joint drives the false-call rate down.

Read →
AREA-SCAN SUITSLINE-SCAN SUITSThe materialDiscrete, indexed partsContinuous web or rollHow it imagesOne snapshot of a fieldOne line at a time, built upMotionPart paused or steppedMaterial moving continuouslyTimingTriggered per partLocked to an encoderBlur riskFrozen by short exposureSmears if speed is not trackedBest forComponents, packs, assembliesFilm, foil, paper, print, coilMatch the camera to whether the material stops. A mismatch fights the line from the first frame.
06VISIONAugust 2026 · 7 min read

The material decides the camera before anything else

Continuous web needs a line-scan camera locked to an encoder; discrete parts suit area-scan. When each is right, and what each demands of the system.

Read →
SUPERVISED CLASSIFIERANOMALY DETECTIONWhat it learnsEach defect from examplesWhat a good part looks likeData it needsMany examples per defectMostly good partsThe rare defectToo few to train onCaught as a deviationA never-seen faultMissed entirelyFlagged as abnormalBest whenDefects are common, well-sampledDefects are rare or unknownThe riskBlind to the unseenTuning normal vs false alarmsYou cannot train on a defect you have barely seen. Model normal instead, and the rare fault stands out.
07AIAugust 2026 · 8 min read

Catching the defect you have barely seen

The defects that matter most are the hardest to train for. How anomaly detection — learning normal from plentiful good parts — catches rare, safety-critical faults a classifier cannot.

Read →
AN IDEAL CHANNELA REAL CHANNELThe signal pathShort, controlled, on-diePackage, board, connector, cableHigh frequenciesPass cleanlyAttenuated, the eye closesDiscontinuitiesNone to speak ofReflect at every junctionJitterNegligibleAccumulates across the pathThe received eyeWide openNearly shut without helpWhat recovers itNothing neededEqualisation, tuned to the channelThe datasheet eye is measured on an ideal channel. The link has to close over the board it actually meets.
08SILICONAugust 2026 · 8 min read

Inside the die it's clean; the trouble starts at the pins

A high-speed link degrades the moment it crosses a package and board — loss, reflections, jitter. How equalisation reopens the eye over a real channel, not an ideal one.

Read →
PASSING TESTS TELL YOUCOVERAGE TELLS YOUWhat was checkedThe cases you wroteHow much of the behaviour was exercisedThe unimagined caseNothingWhether it was reached at allA green regressionThe written tests passSays nothing about the gapsThe escape riskHiddenSurfaced as an uncovered holeStubborn cornersUntouched, unseenReached by directed stimulusHard propertiesUnprovable by simulationClosed by formal proofVerification is showing the untested space has been reached — by stimulus or proof — not declared empty.
09SILICONAugust 2026 · 8 min read

A chip isn't verified because the tests pass

Verified means the untested space is shown empty, not that the regression is green. Functional coverage, constrained-random stimulus, and when only formal can close a property.

Read →
GNSS ON ITS OWNWITH SENSOR FUSIONOpen skyAccurate, metre-class fixPasses through, corrects driftUrban canyonReflections, the fix jumpsMultipath fixes down-weightedTunnel / underpassNo signal, position lostDead-reckons through the gapUnderground / indoorNothing at allInertial holds, drift boundedDense foliageWeak, intermittentWeighted by real reliabilityOn re-acquisitionSnaps back, can jumpCorrects smoothly, no jumpA clear-sky fix is the easy case. The engineering is the tunnel, the canyon and the car park.
10WIRELESSAugust 2026 · 8 min read

When the satellites vanish, the position can't

Satellite positioning fails in tunnels, canyons and indoors. How fusing GNSS with inertial and motion sensors holds an accurate position through the gaps — dead reckoning, multipath rejection, dynamic trust.

Read →
WHAT IT COSTS YOUWHAT IT GIVES YOULicensed (operator band)Lease or an MNO arrangementClean, protected, wide coverageShared / local-licensedRegister, sometimes coordinateDedicated access, low barrierUnlicensedFree, but contendedNo fee, but no protectionLow band, under 1 GHzScarce, narrow channelsRange and wall penetrationMid band, 3 to 4 GHzThe common private choiceBalance of capacity and reachMillimetre-wave, 24 GHz plusShort range, many cellsHuge capacity, line of sightAvailability and rules are country-specific; confirm the local regime before committing to a band.
11WIRELESSAugust 2026 · 8 min read

The band decides the private network you can build

Licensed, shared or unlicensed; low, mid or millimetre-wave — the spectrum you can use sets the coverage, capacity and cost of a private 5G network. Why the regulatory regime is the first design input.

Read →
THE FAULTITS VIBRATION SIGNATUREBearing wearSpalling, roughnessEnergy at bearing-defect frequenciesImbalanceMass off-centreA rise at once-per-revolutionMisalignmentShafts not trueHarmonics, an axial componentMechanical loosenessMounts, worn fitsMany harmonics, broadband riseGear defectA chipped toothSidebands around the mesh frequencyFlow / cavitationPump or fanBroadband, high-frequency energyThe signature is small and buried under normal running; sensing and processing decide whether it survives.
12EMBEDDEDAugust 2026 · 8 min read

A failing machine says what is wrong, in frequency

Predictive maintenance reads early failure in vibration. The signatures a bearing, imbalance or gear defect leaves, why the sensing chain decides if they survive, and separating degradation from normal running.

Read →
WHAT YOU NOTICEWHERE IT ACTUALLY GOESRadio transmitFeels expensiveBrief; rarely the dominant costSleep currentInvisible, ignoredRuns day and night; usually dominatesLocation fix (GNSS)An occasional eventEnergy-heavy; often costly per fixWake frequencyOverlookedEach wake carries fixed overheadPeripheral leakageAssumed powered downA floating pin can drain for monthsCold temperatureNot consideredCapacity falls; current pulses sagA few microamps of avoidable sleep current outweighs a year of transmissions.
13EMBEDDEDAugust 2026 · 8 min read

The battery is spent where you don't look

Battery life in a field device is set by sleep current and location, not the radio. How to build an energy budget, hunt quiescent draw to microamps, and prove multi-year life without a multi-year test.

Read →
THE GAPWHAT HANDLES ITUrban canyon / tunnelGNSS drops, cellular fadesDead-reckoning and store-and-forwardRemote routeNo cellular coverage at allBuffer on device, satellite fallbackBorder crossingRoaming and regulatory changeMulti-network roaming, eUICCUnderground car parkDeep indoor, no signalHold on device, sync on exitCongested networkLatency and packet lossPrioritise, compress, batchIgnition offThe device must still persistLow-power keep-alive, wake on eventThe platform assumes a live link. The device’s real job is to behave correctly, and keep the data correct, when there isn’t one.
14CONNECTIVITYAugust 2026 · 8 min read

A fleet is only as connected as its worst dead zone

Fleet telematics is not the dashboard — it is keeping data flowing and correct across the coverage a vehicle loses: tunnels, remote routes, borders. Bearers, store-and-forward, dead-reckoning and roaming.

Read →
SIGNALWHAT IT SHOWSEye closure (PERCLOS)Fraction of time the eyes are closedThe strongest early indicatorBlink dynamicsSlow, long, heavy blinksMicro-sleep onsetHead poseNodding and postural driftLoss of alert controlGaze directionEyes leaving the roadDistraction, distinct from fatigueSteering behaviourMicro-corrections and driftVehicle-side corroborationFacial expressionYawning and slackeningSupporting evidence, never aloneNo single signal is enough. Fatigue is inferred from several, fused, before the alarm is worth a driver’s trust.
15AIAugust 2026 · 8 min read

A tired driver looks fine until the moment they don't

Detecting driver fatigue from eye closure, head pose, gaze and steering — why it runs in the cab, works in darkness, tells fatigue from distraction, and lives or dies on its false-alarm rate.

Read →
STAGETHE CATCHPhysical tapOBD-II port or a CAN wire-inRead without disturbing the busDecodeStandard PIDs or a DBC mapSignals differ by make and modelSample and filterPick the signals that matterNot every frame is worth sendingLocal rulesHarsh-brake, idle, geofenceDetected on the device, not the cloudTimestamp and bufferA trusted clock, local storeSurvive the coverage dropoutsUplinkForward events, not raw framesThe bus is a firehose; send the signalThe bus is a firehose. The engineering is deciding, on the device, what is an event and what is noise — and never writing back to the bus.
16EMBEDDEDAugust 2026 · 7 min read

The vehicle already knows everything; telematics is getting it out

Tapping a vehicle's buses safely over OBD-II or CAN, decoding signals that differ by make, detecting events on the device, and shipping only what matters — without disturbing the bus or opening a security hole.

Read →
SIGNAL STRENGTH (RSSI)CHANNEL SOUNDINGWhat it measuresReceived power, mapped to a guessRadio phase and round-trip timeAccuracyMetres, and it driftsCentimetre scale in good conditionsMultipathReflections corrupt the estimateResolved across many frequenciesSecurityAmplify the signal to fake nearnessTwo-way timing resists relay attacksMethodOne number, no physics of distancePhase-based ranging plus round-trip timeRolesAny receiver, passivelyA connected initiator and reflector pairRSSI infers range from how loud a signal arrives, which reflections and an attacker can both change. Channel Sounding measures the distance itself.
17CONNECTIVITYAugust 2026 · 8 min read

Bluetooth Channel Sounding: measuring distance, not guessing it

Signal strength was never a distance sensor. Channel Sounding, added in Core 6.0, measures true physical distance to centimetre scale with phase-based ranging and round-trip time — accurate, and hard to spoof.

Read →
WHAT IT ISREUSE OR RE-TESTController — PHY and link layerThe radio and the lowest layersInherit the module’s qualified designHost — L2CAP, GATT, GAP, SMThe protocol stack above the radioReuse a qualified stack, re-test if changedProfiles and servicesLE Audio, HID, custom GATT servicesQualify what you actually addICS / IXITThe features you declare you supportPer product; it decides which tests applyConformance testingTest cases drawn from the reference listIn-house if covered; a lab for new RFDeclaration and listingThe end product recorded with the SIGRequired before you ship or use the marksYou qualify subsystems once and reuse them. A new Core feature — Channel Sounding, LE Audio — puts more of the stack back into 'new', which is where the effort goes.
18CONNECTIVITYAugust 2026 · 8 min read

Bluetooth qualification, and what it takes to ship

You cannot ship a product using the Bluetooth marks until it is qualified and listed. You inherit most of it from a pre-qualified module; the effort is proportional to what you add — a profile, new RF, or a Core feature like Channel Sounding.

Read →
WHAT THE STAGE DOESWHY IT MATTERSGround-plane calibrationLevels and segments the ground, continuouslySeparates track bed from real objectsRail-line detectionFinds the rails in the point cloudDefines where the train will actually beDynamic area of interestTrack corridor plus a margin, updated liveOnly the danger zone raises an alarmObject classificationSize and density — person, animal, debrisA small animal is not a person on the lineTracking and forecastPose, velocity and predicted pathIs it moving onto the track, or away?Alert and latencyThreshold breach triggers the train alertDetection-to-alert time is the safety metricLiDAR gives range and shape in darkness, glare and weather. The pipeline turns the point cloud into one decision: alert, or not.
19VISIONAugust 2026 · 9 min read

Rail intrusion detection with LiDAR, from point cloud to alert

A LiDAR returns millions of points; a safety system has to return one bit. Ground-plane calibration, dynamic rail detection, classification by size and density, tracking, and a train alert bounded by response time.

Read →
SHOWS UP ASFIXED INFloorplan shape & pin sidesGlobal-route overflow at edgesFloorplanHigh placement utilisationShorts and spacing DRCs in hotspotsPlacement / target densityStandard-cell pin densityPin-access and via failuresCell padding, boundary cellsHigh-fanout and clock netsLocal congestion clumpsRTL restructure, CTSMacro halos and channelsDetours, unroutable channelsFloorplan, keep-outsLayer resources and blockagesAntenna and layer starvationLayer / NDR assignmentA router places wires in the space placement left it. Most congestion is closed before detailed routing.
20SILICONAugust 2026 · 9 min read

Routing congestion, and why place-and-route fails

A block can meet timing on paper and still be unroutable. When routing fails, the cause sits in floorplan and placement — not the router.

Read →
VCDeVCD (EXTENDED)RecordsValue changes with timestamps+ drive strength and directionBidirectional portsDirection inferredExplicitly capturedStandardIEEE 1364, ASCIIIEEE 1364 extended, ASCIITypical useDebug, toggle activity, powerATE/test, drive conflicts, bidir powerSize, full chipLarge — tens to hundreds of GBLarger stillFeedsSAIF → power toolSAIF/power plus test vectorsBoth are plain-text activity dumps. eVCD adds strength and direction for bidirectional and test cases.
21SILICONAugust 2026 · 8 min read

VCD and eVCD: managing activity, and what it drives

Value Change Dump and its extended form are the currency of power analysis — but a full-chip dump reaches hundreds of GB. What you dump decides whether numbers arrive on time.

Read →
01Two carriersf₁ and f₂, high power02Metallic junctionoxide, low torque, dissimilarmetals03Weak non-linearitythe joint rectifies04Mixing products2f₁−f₂ and 2f₂−f₁05Lands in receive bandcannot be filtered06Site desensitisesuplink coverage shrinksThird-order products rise about 3 dB for every 1 dB of carrier increase, so adding a carrier can expose PIM that was always present.
22WIRELESSAugust 2026 · 11 min read

Passive intermodulation: the fault with no active component

Two carriers meet a corroded joint, the joint behaves like a diode, and the site desensitises. Nothing in the electronics has failed.

Read →
ADDS LATENCYTHE LEVERCapture & exposureOne frame plus exposure timeHigher frame rateEncodeCodec pipeline and GOP structureLow-latency encoder, short GOPRadio transportAir time plus retransmissionLink margin, less ARQJitter bufferSmoothing delay before decodeSmaller buffer, accept some lossDecodeDecoder pipeline depthHardware decodeDisplayPanel and compositorDirect scan-outA radio quoted at 20 ms can sit inside a 400 ms system. Glass-to-glass is the only figure that matters.
23WIRELESSAugust 2026 · 9 min read

Latency in a drone video link, measured honestly

A radio quoted at twenty milliseconds can sit inside a system that takes four hundred. Glass to glass is the only figure that matters.

Read →
SET BYDECIDESSensor resolutionPixels needed across the defectWhether the defect is resolvableMagnification / FOVPart size and placement spreadField of view and lens choiceFocal lengthWorking distance availableStandoff and packagingAperturePart height variationDepth of field in focusExposure / strobeLine speed and motionMotion blur at the worst caseIllumination geometryHow the defect scatters lightWhether it shows contrast at allImaging first, algorithm second. A defect with no contrast under the light carries no information to find.
24VISIONAugust 2026 · 11 min read

Choosing the camera, the lens and the trigger

Most inspection systems that fail were decided at the lens, not at the model. Resolution runs backwards from the defect.

Read →
THE SPECTRUMWHICH ONERangeTouch, to metres, to tensNFC cm · RFID m · UWB farPositionNone, to sub-metreUWB for precise locationPowerPassive to activeRFID/NFC passive · UWB activeCost per tagCents to dollarsPassive RFID is cheapestData exchangeBit of ID, to secure pairingNFC for a secure tapBest questionPresence · tap · locatePick by the question, not the tagThey are not competitors so much as answers to different questions: is it here, is it authentic, exactly where is it.
25CONNECTIVITYAugust 2026 · 10 min read

RFID, NFC and UWB: what each one is actually for

Three technologies answering three different questions. Choosing on the wrong axis is the expensive error here.

Read →
TWO-WAY RANGINGTDoAWhat it measuresRound-trip time of flightTime difference at anchorsClock syncNot needed between devicesAnchors must be synchronisedInfrastructureTwo radios talk to each otherMany synchronised anchorsTag scalingPer-pair exchange, slowerOne blink serves all anchorsTag powerHigher, tag transmits and listensLower, tag only blinksBest fitFew tags, direct distanceMany tags, real-time locationTwo-way ranging needs no synchronised infrastructure but scales poorly; TDoA needs synced anchors but scales to many tags. PDoA adds angle.
26CONNECTIVITYAugust 2026 · 10 min read

How UWB actually measures distance

Thirty centimetres per nanosecond, so sub-metre accuracy is a sub-three-nanosecond timing problem. Everything hard follows from that.

Read →
WHAT IT CARRIESWHY IT IS BEAMEDSS/PBCH blockPSS, SSS and the PBCHThe UE's first anchor to the cellBeam sweepEach SSB on a different beamOne beam cannot cover the sectorSSB burst setA group of beams in a windowSweeps the whole coverage areaPeriodicityRepeats on a set intervalLets idle UEs find and measure itBeam indexRead from the time positionTells the UE which beam it heardCoverage vs countMore beams, narrower eachRange traded against sweep timeAt higher bands a cell cannot shout in all directions at once, so it sweeps a set of SSB beams in time.
27WIRELESSAugust 2026 · 11 min read

How 5G NR redefined the cell

A cell is now a synchronisation block and an identity, not a transmitter — which is why moving across a building need not generate a handover.

Read →
STARDAISY-CHAINCablingHome-run to each unitOne cable threaded unit to unitLatencyUniform, one hopAccumulates along the chainA breakIsolated to one unitCuts everything downstreamTiming / syncDistributed from the hubPassed through each nodeInstall costMore cable, more portsLess cable, fewer portsScaling limitHub port countHop count and delay budgetDaisy-chaining saves cable and ports; it pays for it in accumulated latency and a single point that breaks the rest.
28WIRELESSAugust 2026 · 9 min read

Radio cascading: how deep can you daisy-chain radio units?

Bandwidth, timing accumulation and failure domain each bound the chain. The tightest one wins, and it is usually bandwidth.

Read →
CENTRAL UPFEDGE UPFRound-trip latencyHigher — traffic hairpins inwardLow — served near the userLocal breakoutNo, everything goes centralYes, traffic exits locallyBackhaul loadHeavy, all user trafficOffloaded at the edgeEdge compute (MEC)Far from the applicationCo-located with the UPFOperationsSimple, few sitesDistributed, many sites to runBest fitDefault mobile broadbandURLLC and local-data use casesPushing the UPF to the edge cuts latency and backhaul; it buys that with the cost of operating many distributed sites.
29WIRELESSAugust 2026 · 10 min read

The user plane function, and where it should actually sit

Central, edge or cell site — the placement sets the latency floor. It is also the anchor where 5G and Wi-Fi converge on one session.

Read →
SPREADS THE SYMBOLROOT CAUSEPhase noiseRotational smearing of pointsOscillator and PLL qualityIQ imbalanceSkewed, distorted constellationModulator calibrationPA compressionOuter points pulled inwardBackoff or digital pre-distortionFrequency errorWhole constellation rotatesLocal oscillator accuracyThermal noiseFuzzy cloud around each pointSNR and the gain planTiming errorSmearing along the axesSample-clock accuracyEVM collapses many different faults into one figure — which is why a bad number still needs the constellation to diagnose.
30WIRELESSAugust 2026 · 9 min read

EVM: what it measures, and what it hides

Six mechanisms degrade a constellation, they add in power, and the largest dominates — so decompose before optimising.

Read →
WHAT IT CAUSESTECHNIQUEPA nonlinearitySpectral regrowth into neighboursDigital pre-distortionHigh PAPREfficiency lost to backoffCrest-factor reductionMemory effectsAsymmetric, bandwidth-dependent distortionMemory-polynomial DPDLow efficiencyHeat and wasted DC powerDoherty PA plus DPDACLR limitsLeakage into adjacent channelsDPD to specificationIn-band errorEVM eats the modulation marginLinearisation budgetDPD and CFR exist to let an efficient, nonlinear PA meet a linear spectral mask — efficiency and cleanliness at once.
31WIRELESSAugust 2026 · 10 min read

Linearisation in 5G: why the amplifier decides the power bill

CFR and DPD exist because OFDM peaks sit where amplifiers are least efficient. That efficiency is a recurring operating cost.

Read →
MOCNMORANRadio hardwareShared across operatorsShared across operatorsSpectrumShared, pooled carriersSeparate carrier per operatorCore networkSeparate per operatorSeparate per operatorBroadcastMultiple PLMN IDs on one carrierPer-operator carrier and IDsRegulatory fitWhere pooling is allowedWhere spectrum must stay separateNeutral hostHost owns and shares the RANSame, with per-operator carriersMOCN pools the spectrum; MORAN keeps it separate. A neutral host runs either on behalf of the operators it carries.
32WIRELESSAugust 2026 · 9 min read

MOCN, MORAN and neutral host: three ways to share a radio network

What each model shares, the broadcast identity limit that caps tenants, and why sharing fails on operations rather than radio.

Read →
WHAT IT DOESWHERE IT SITSSecure tunnelIPsec over the untrusted accessBetween UE and N3IWFNWu interfaceCarries NAS and user trafficUE to N3IWF over Wi-FiControl planeRelays NAS to the coreN2 to the AMFUser planeForwards data to the coreN3 to the UPFAuthentication5G AKA carried over EAPUE, N3IWF, AUSFTypical useWi-Fi calling, fixed-wireless offloadAlongside the 3GPP accessUntrusted non-3GPP access reaches the 5G core only through an IPsec tunnel to the N3IWF — the core sees a standard N2/N3.
33WIRELESSAugust 2026 · 8 min read

N3IWF: bringing Wi-Fi into the 5G core

Reaching the 5G core over a network the operator does not own — and where that is cheaper than building coverage.

Read →
LTE ANCHORNR LEGCarries signallingYes — RRC and mobility controlUser plane, split from the anchorCoverageWide, on a low or mid bandNarrower, on a higher bandPrimary roleControl and the coverage floorAdded throughput and capacityIf the other dropsStays up and holds the callDropped first; anchor survivesBand choiceExisting LTE FDD/TDDMid-band or mmWave NRCombining methodDual connectivity, not carrier aggregationSame — two schedulers, one UEIn 5G non-standalone, the LTE anchor owns control and mobility; the NR leg is added capacity on top.
34WIRELESSAugust 2026 · 8 min read

Anchor bands and EN-DC: why 5G still needs LTE

In non-standalone deployment 5G coverage cannot exceed the LTE anchor, which explains much of what users actually experience.

Read →
NB-IoTLTE-MData rateLow — small, infrequent readsHigher — firmware and richer payloadsMobilityBest for fixed metersHandles moving assetsVoiceNot supportedVoLTE possiblePowerLowest, deep sleep friendlyLow, but higher than NB-IoTCoverageDeepest indoor penetrationGood, not as deepBest fitStatic utility metersMobile or update-heavy devicesBoth are low-power WAN. NB-IoT trades rate for reach and battery; LTE-M trades battery for rate and mobility.
35CONNECTIVITYAugust 2026 · 9 min read

Choosing cellular for smart metering: NB-IoT, LTE-M and RedCap

Decided by coverage in the worst basement in the estate, not by throughput — and by whether the network still exists in year twelve.

Read →
WHAT IT NEEDSIF IT FAILSAlways-on linkDual bearer with failoverA dropped link is a lost salePayment securitySecure element, minimal PCI scopeA breach ends the deploymentRecoveryFast boot, watchdog, rollbackA frozen till blocks the queueUpdatesOTA across the whole estateA bad update bricks thousandsPeripheralsPrinter, scanner, card readerOne dead peripheral halts checkoutEstate scaleFleet monitoring and controlStore-by-store does not scaleThe hard part is not the terminal; it is running thousands of them, updated and online, without a truck roll.
36CONNECTIVITYAugust 2026 · 8 min read

Connected point of sale: where the engineering actually is

A payment terminal is a security product that happens to take payments. Tamper response and certification govern the schedule.

Read →
WHAT YOU LEARNFITTING TECHIs it here?Presence at a gate or portalPassive RFIDWhich zone?Room or area it is inBLE beaconsExactly where?Sub-metre positionUWBIn transit?Location out of coverageCellular and NTNTampered?Opened or breachedSensor tagChain of custody?Who handled it, and whenEvent logItem-level tracking is a stack of questions, not one technology. Each granularity has a different fitting radio.
37CONNECTIVITYAugust 2026 · 11 min read

Item-level tracking for high-value goods in distribution logistics

Density, custody handovers and tamper evidence — and why DECT NR+, with no operator and no base station, changes the arithmetic.

Read →
STAY ON FPGACONVERT TO ASICAnnual volumeLow or uncertainHigh and sustainedRequirementsStill movingStable and validatedBinding constraintTime to marketUnit cost, power or board areaProduct life remainingShortLong enough to amortiseField updatesNeeded after deploymentNot requiredIf requirements are still moving, conversion is premature regardless of the volume arithmetic.
38SILICONAugust 2026 · 9 min read

When does FPGA-to-ASIC conversion make business sense?

The volume arithmetic, the thresholds, the hidden costs estimates miss, and the cases where the honest answer is no.

Read →
01Encoder triggerlocked to part position02Captureexposure inside motionbudget03Inferenceworst-case latencymeasured04Verdictinside the PLC window05Part trackingshift register acrossstops06Rejectdownstream actuation07MES recordverdict on the batchA correct verdict applied to the wrong part is worse than no inspection at all.
39AIAugust 2026 · 8 min read

How AI visual inspection integrates with PLC and MES

Trigger and encoder, verdict timing, part tracking, reject actuation, the fault protocol and traceability.

Read →
CELLULARSATELLITE (NTN)Energy per messageLowSubstantially higherCost per messageLowHigherCoverageWhere operators reachEffectively everywhereUsed forRoutine reportingAlarms and continuityPayloadFullReduced deliberatelyBoth transports reconcile into one asset history — the gap is where loss and delay happen.
40CONNECTIVITYAugust 2026 · 7 min read

Cellular-first tracking with NTN satellite fallback

How coverage continuity is architected, what escalation policy should look like, and what a satellite message actually costs.

Read →
RELATIVE RANGENFC≈10 cm · no batteryRFID UHF1–12 m · no batteryUWB10–50 m · sub-metre, monthsBLE10–100 m · 1–5 yearsWi-Fi30–100 m · mains typicalLoRaWANkm · 5–10 yearsCellularwide area · weeks–yearsSatelliteglobal · highest per messageRange, power, accuracy and infrastructure trade against each other. Nothing wins on all four.
41CONNECTIVITYAugust 2026 · 8 min read

Choosing a wireless link for asset tracking

NFC, RFID, BLE, UWB, LoRaWAN, cellular and satellite compared on range, accuracy, battery life and infrastructure.

Read →
THE UNIT, AND THE ARRAY Hash corepipeline + rounds Replicated array — limited by power and heat, not logic THE OBJECTIVE min. joules / hash min. cost / hash Everything else is subordinate to these two.
42SILICONAugust 2026 · 7 min read

Designing a crypto-mining ASIC: the economics are in the floorplan

A mining chip competes on energy per hash and cost per hash. Why the win is in the micro-architecture, the power delivery and the floorplan — not the algorithm.

Read →
MATURE NODEADVANCED NODEMask and NRE costLowVery highAnalog performanceBetter headroom and matchingConstrainedEmbedded non-volatile memoryWidely availableOften unavailableDigital density and speedLowerThe reason to go thereLeakageSmallFirst-order below 28 nmSupply horizonLongShorter, faster-movingNode is an architecture decision. Analog-heavy and long-lifecycle parts frequently belong on mature.
43SILICONAugust 2026 · 7 min read

Process node selection for mixed-signal and secure silicon

Why 180 nm is frequently the right answer, and when an advanced node genuinely earns its mask cost.

Read →
SINGLE IMAGEDUAL IMAGE (A/B)Flash costLowestTwo full application slotsPower loss mid-writeBricks the unitRunning image untouchedRollbackNot possibleAtomic switch backBad release exposureThe installed baseA canary groupRetrofittableNoNo — decided at part selectionPartition layout, signing and rollback are fixed before the first unit ships.
44EMBEDDEDAugust 2026 · 8 min read

Designing over-the-air update before you ship

Partition layout, bootloader design, signing and staged rollout — decisions that cannot be retrofitted into a deployed fleet.

Read →
01Constraintstated as a number02Real datayour parts, your line03Sensing trialseveral configurations04Baseline modelnot a product05Measured resultagainst the constraint06Recommendationincluding: do not proceedA study concluding it will not work as posed usually saves a multiple of its own cost.
45AIAugust 2026 · 6 min read

What an AI feasibility study actually produces

Two to six weeks, real parts, real conditions — and a written answer that is sometimes that it will not work as posed.

Read →

Filtering is a convenience. Every article is listed above and none is hidden behind it — if JavaScript is off, you see all of them.

ALSO HERE

Beyond the articles.

DOWNLOADS

White papers

Longer technical documents, available on request. Short form, no drip sequence afterwards.

REFERENCE

Glossary

Thirty-one plain definitions of the terms used across this site, written by the people doing the work.

EVIDENCE

Case studies

Five programmes described at property and outcome level, with the hard parts named.

SUBSCRIBE

Engineering notes, monthly at most.

New articles, reference architectures and application notes as they are published. No product announcements, no event invitations, and no more than one email a month.

Unsubscribe from any email in one click. Your address is not shared outside Faststream Technologies and is not used for anything except this list.

EDITORIAL POLICY

What gets published here.

Articles are written by engineers who have done the work, and they include the constraints and failure modes rather than only the successes. Where a technique has a boundary, the boundary is stated.

Nothing here contains customer-identifying material. Figures are published with units and with the configuration they were measured under, or they are not published.

There is an RSS feed at /knowledge/feed.xml if you would rather not receive email.

COMMON QUESTIONS

What engineers ask before they call.

01

Are white papers gated?

Selected longer documents require a short form — name, work email and organisation. Everything in the Knowledge Center itself is open, because gating an explanation defeats its purpose.

02

How often is the Knowledge Center updated?

Irregularly, when there is something worth writing. The newsletter goes out no more than once a month, and there is an RSS feed for anyone who prefers not to give an email address.

03

Can these articles be shared or quoted?

Yes. Share links are on every article. Quoting with attribution and a link back is welcome.

KEEP READING

Related work.

CASE STUDIES

Case studies

Deployment stories with outcomes.

PRODUCTS

Products

What can be licensed and bought.

COMPANY

Engineering excellence

How Faststream works.

BUILD WITH FASTSTREAM

Bring us the difficult part.

Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.

Start a projectRequest an architecture review