Duplexing: FDD versus TDD

Every lesson in this module so far has answered the same question: how do you fit many users into one piece of spectrum? Frequency (M10-L1), time (M10-L1), code (M10-L1), subcarriers (M10-L2), space (M10-L3). This lesson answers a different question that looks deceptively similar, and the confusion between the two is the single most common muddle in wireless. A phone call is not one flow of information; it is two, and they run at the same time in opposite directions. Deciding how those two directions share the radio is duplexing, and there are only two answers in commercial use: give them separate frequencies, or give them separate instants.

Two Directions, One Link

Name the directions first, because the rest of the lesson leans on them. The downlink carries information from the base station or access point to the user; the uplink carries it back. They are not equal in any respect. M9-L4 showed that they are not equal in power — a macro base station transmits at 46 dBm and a handset at 23 dBm, a 23 dB advantage to the downlink which the base station’s better noise figure only claws 5 dB back from, leaving the uplink as the binding constraint by 18.0 dB. And they are not equal in demand either: a user streaming video, loading a page or scrolling a feed pulls far more bits down than up.

So the duplexing question is really: given one asymmetric link with two unequal directions, what resource do you spend to keep them from destroying each other? A transmitter beside a receiver is the loudest thing that receiver will ever encounter, by an enormous margin. Separating the two is not a nicety; it is the whole engineering problem, and the two commercial answers spend two different resources on it — FDD spends spectrum, TDD spends time.

Duplexing Is Not Multiple Access

Because the names rhyme, FDD gets read as a synonym for FDMA and TDD as a synonym for TDMA. They are independent choices, and the cleanest proof is that every combination has actually been deployed. Multiple access divides a channel among users; duplexing divides it between the two directions of one user’s link. A system makes both choices, separately:

Read the GSM row and the LTE TDD row together and the point is settled. GSM is the archetype of TDMA and it duplexes by frequency; LTE TDD divides users by frequency and duplexes by time. If the two concepts were the same thing, neither row could exist. The reliable way to keep them apart is to ask what is on the other side of the divide: another user means multiple access, the same user talking the other way means duplexing.

Half Duplex, Full Duplex, and the One Nobody Has Deployed

A half-duplex link carries one direction at a time and the users take turns explicitly — a walkie-talkie, where you press to talk and hear nothing while you do. A full-duplex link carries both directions at once, which is what a telephone conversation needs and what every cellular system provides. Here is the sentence that matters: FDD and TDD both deliver a full-duplex service, and they differ only in how. FDD is genuinely simultaneous. TDD alternates so quickly — typically switching every 5 ms or faster, against a speech frame of 20 ms — that the conversation is seamless, so the user experiences full duplex even though the radio is half duplex at every instant. TDD is best described as fast-alternating half duplex packaged as full duplex.

There is a third possibility, and it is worth being honest about its status. In-band full duplex would transmit and receive on the same frequency at the same time, doubling spectral efficiency at a stroke. The obstacle is arithmetic. Your own transmitter is right beside your own receiver, so its leakage arrives at roughly the power you sent while the wanted signal arrives at roughly the receiver sensitivity:

So treat in-band full duplex as a research topic, not an option. When a specification says “full duplex” it means the service is bidirectional, and the physical layer underneath is FDD or TDD. One further wrinkle in the terminology, worth knowing because it appears on datasheets: some FDD devices are half-duplex FDD, meaning they have the paired bands but never use both at once, which lets them omit the duplexer and use a switch instead. Cellular IoT categories do exactly this to save cost.

• Half duplex — one direction at a time, explicitly. Walkie-talkie, and cellular IoT devices that skip the duplexer
• FDD — both directions always on, on two different frequencies. Truly simultaneous, and needs paired spectrum
• TDD — both directions on one frequency, alternating fast enough that the user cannot tell. Needs one unpaired block
• In-band full duplex — one frequency, one instant, both directions. Needs 100+ dB of self-interference cancellation and is not deployed

FDD: Buy Two Bands and a Filter

Frequency-division duplex gives the uplink one band and the downlink another, permanently. Both radios transmit continuously and receive continuously, and the isolation between them is bought with filters. The spectrum has to be allocated in pairs, which is why regulators speak of paired and unpaired spectrum, and why the pairing is fixed in the band plan for the lifetime of the band. Work through LTE Band 1, the classic 2 GHz pair, with all the arithmetic on the page:

• Uplink: 1920–1980 MHz, so the uplink band is 1980 − 1920 = 60 MHz wide
• Downlink: 2110–2170 MHz, so the downlink band is 2170 − 2110 = 60 MHz wide — the pair is symmetric, and that symmetry is the whole problem discussed below
• Duplex spacing is the offset between the two band edges, 2110 − 1920 = 190 MHz. A handset tunes its receiver exactly 190 MHz above its transmitter, always
• Duplex gap is the empty spectrum between the two bands: 2110 − 1980 = 130 MHz. Check it against the other two figures: 190 − 60 = 130 MHz ✓
• That gap is 130/2045 = 6.4% of the centre frequency, which is what makes Band 1’s filters comparatively easy. Band 3 is much tighter: 1710–1785 up, 1805–1880 down, so 75 MHz each, 95 MHz spacing, and a gap of only 1805 − 1785 = 20 MHz

The duplexer is the hidden cost of FDD

The component that makes FDD work is the duplexer: a pair of sharp filters sharing one antenna port, one passing the transmit band towards the antenna and blocking it from the receiver, the other passing the receive band to the receiver. It typically provides 50–60 dB of transmit-to-receive isolation and costs about 1–2 dB of insertion loss in each direction — and M9-L4 priced exactly what a decibel of loss is worth, so this is a real line in the budget, not a footnote.

Here is the consequence that shapes handset design: a duplexer is tuned to one band pair, so a phone needs one duplexer per FDD band it supports. A modern handset supports tens of bands, and the resulting bank of filters is a genuine driver of board area, bill of materials and antenna-path loss. TDD needs no duplexer at all — only a switch, which is far smaller, cheaper and broader in bandwidth. That single hardware fact is a large part of why the newer, wider bands went to TDD.

TDD: Buy One Band and a Stopwatch

Time-division duplex gives both directions the whole band and alternates between them. One unpaired block is enough, and the transmitter and receiver never operate at the same instant, so no filter has to separate them. What TDD pays instead is a guard period: an interval of deliberate silence at each downlink-to-uplink switch. It has to absorb two things, and the larger one is geometry. Radio travels 3×10⁸ m/s, which is a convenient 300 metres per microsecond, so a user at the cell edge hears the downlink late and its uplink burst arrives back late again:

• 1 km cell: 2 × 1000 / (3×10⁸) = 6.67 µs. At a 5 ms switching period that is 6.67/5000 = 0.13% overhead — free
• 10 km cell: 2 × 10 000 / (3×10⁸) = 6.667×10⁻⁵ s = 66.7 µs, or 66.7/5000 = 1.33% — still cheap. Cross-check with the 300 m/µs rule: 20 km ÷ 300 m per µs = 66.7 µs ✓
• 30 km cell: 200 µs, which is 4.0% of every switching period spent transmitting nothing
• 100 km cell: 666.7 µs, or 13.3% — and this is why very large TDD cells are unattractive rather than impossible
• A geostationary satellite: 2 × 35 786 km / (3×10⁸) = 239 ms of round trip, roughly fifty switching periods. TDD framing simply does not apply, which is one reason satellite links are overwhelmingly FDD

LTE TDD’s special subframe, and the flexibility payoff

LTE TDD implements the guard inside a dedicated special subframe, split into a downlink part, the guard period, and an uplink part. The guard is configurable in whole OFDM symbols, and one symbol with the normal cyclic prefix (M10-L2) is 1 ms / 14 = 71.4 µs, which covers a radius of 71.4 µs × 300 m/µs ÷ 2 = 10.7 km. Ten symbols of guard, the largest setting, reach 714 µs and therefore 107 km — at the cost of most of a subframe. The overhead is not hidden in TDD; it is a number an operator dials in against the cell size it actually built.

In exchange, the downlink-to-uplink split becomes a configuration item rather than a property of the band. LTE TDD defines seven uplink-downlink configurations; configuration 2, with a 5 ms switching period, allocates three downlink subframes to one uplink subframe — a 3:1 ratio — while configuration 1 is 2:2 and configuration 0 is uplink-heavy. 5G NR generalises this further, letting the pattern be signalled and changed. An FDD band cannot do any of this: its 60 MHz up and 60 MHz down were decided by a regulator before the traffic existed.

The Trade-offs, Each With Its Reason

A bare list of advantages is useless, because every item on it is a consequence of something already established. Four trade-offs matter, and each one follows from a fact from an earlier lesson.

Traffic asymmetry — TDD’s biggest win

Real internet traffic is lopsided, often three or four times more downlink than uplink. TDD can allocate its slots to match: ask for 4:1 and you get 4:1. FDD cannot, because the split is 50:50 by construction — the two bands are equal and permanently assigned. Put a number on the loss. Let the demand ratio be r downlink to 1 uplink, and normalise the total resource to 1. FDD holds 0.5 in each direction, so it can serve traffic only up to the point where the busier direction runs out, and the quieter direction’s leftover cannot be lent across:

TDD pays only its guard overhead, which the previous section priced at 1.33% for a 10 km cell — so against 4:1 traffic, TDD keeps about 98.7% of the resource and FDD keeps 62.5%. That is the largest single number in this lesson, and it explains why every band opened for mobile broadband since about 2010 has been unpaired. Note the honest converse: at symmetric traffic, r = 1, FDD wastes nothing and TDD still pays its guard, so FDD wins outright on symmetric loads. Voice-only networks, which are symmetric by nature, were built on FDD for exactly this reason.

Channel reciprocity — why massive MIMO prefers TDD

In TDD both directions use the same frequency, so they see the same physical channel. That is reciprocity, and it is enormously valuable: a base station that measures the uplink knows the downlink, without the user having to describe it. M10-L3 needs precisely this. Beamforming and spatial multiplexing require channel knowledge at the transmitter, and for a 64-element array the amount of information to feed back is large enough that explicit feedback becomes the system’s bottleneck. In TDD it is free. In FDD the uplink and downlink are 190 MHz apart in Band 1 — far more than any coherence bandwidth (M8-L3 gave 4 MHz indoors and 200 kHz in a city) — so the two channels are unrelated and the user must measure and report the downlink explicitly.

So reciprocity is not free of conditions, and the honest summary is that TDD reciprocity degrades with speed. A system that sounds the channel every slot rather than every switching period does better, and that is exactly what massive MIMO deployments do; but the trend is unavoidable, and it is why the largest antenna arrays are aimed at dense, slow-moving urban traffic rather than at motorways. Reciprocity also assumes the transmit and receive hardware chains are themselves matched, which they are not by nature — real base stations run a calibration loop to make them so.

Range, synchronisation and interference

FDD transmits continuously, which suits long range in two ways. First, its transmit duty cycle is 100%, whereas a TDD cell transmitting downlink 75% of the time must either accept 10 log₁₀(0.75) = −1.25 dB of average power or raise its peak power to compensate, and peak power is what the amplifier and the regulator both limit. Second, FDD has no guard period, so cell size costs it nothing at all — the 13.3% penalty a 100 km TDD cell pays does not exist in FDD, and the 239 ms of a satellite round trip is simply irrelevant to it.

TDD carries a requirement FDD does not have at all: network-wide synchronisation. If two neighbouring TDD cells on the same channel are not aligned, one will be transmitting downlink at full base-station power while the other is trying to receive a 23 dBm handset. M9-L3 named this co-channel interference and M9-L4 priced the asymmetry — here the interfering signal can be tens of decibels above the wanted one, arriving from a base station with line of sight rather than from a shadowed user. It is not a degradation, it is an outage. TDD networks therefore synchronise to a common clock, usually GNSS-derived, and neighbouring operators in the same band must agree on a common frame structure. FDD operators need none of this coordination, which is one reason FDD is administratively simpler in fragmented spectrum.

Hardware

TDD needs a fast transmit/receive switch where FDD needs a duplexer, and a switch is cheaper, smaller and broadband where a duplexer is expensive, bulky and per-band. Against that, TDD needs a receiver that recovers from its own transmitter’s blast within microseconds, and it needs the synchronisation infrastructure above. FDD needs neither, but pays 1–2 dB of duplexer insertion loss in both directions and one filter pair per supported band. Summarised in one line: FDD spends spectrum and silicon area to avoid coordination; TDD spends coordination and a little time to avoid spectrum and silicon area.

Where Each Is Used

The split in real deployments follows the trade-offs almost exactly, and it also follows history: the paired bands were allocated when traffic was symmetric voice, and the unpaired bands were allocated when it was asymmetric data.

5G NR supports both, and specifies its bands accordingly — the n1, n3 and n8 numbers are the paired FDD bands inherited from LTE, while n41 upward are unpaired TDD. The pattern is clear enough to state as a rule of thumb: below 3 GHz expect FDD, above it expect TDD, and in mmWave expect TDD without exception.

What Module 10 Built

This lesson closes the module, and the module has a shape worth naming. Every lesson in it answered one question — how is a finite piece of spectrum shared? — and together they enumerate the answers:

• By frequency — FDMA, one narrow slice per user, with guard bands between them (M10-L1)
• By time — TDMA, one slot per user, with guard times between them (M10-L1)
• By code — CDMA, everyone at once, separated by orthogonal sequences (M10-L1)
• By orthogonal subcarrier — OFDM and OFDMA, which turned the multipath of M8-L3 from a hazard into a bookkeeping problem, and made frequency-domain sharing fine-grained (M10-L2)
• By space — MIMO, where several antennas send parallel streams through the same frequency at the same time, and spatial multiplexing buys capacity that none of the other four could (M10-L3)

And this lesson added the dimension that sits across all five: however you split the channel among users, you must still split it between the two directions of each user’s link, by frequency or by time. That completes the toolkit. Module 11 is where real standards pick their combination — WiFi choosing contention plus TDD, LTE choosing OFDMA plus either duplexing, 5G NR choosing OFDMA plus massive MIMO plus TDD in the new bands — and every one of those choices is now a choice you can read rather than memorise.