Every generation before this one was designed around a single dominant service. 2G was designed for a voice call, 3G for a voice call with a data channel bolted alongside it, and 4G LTE for a packet pipe to a smartphone (M11-L2). Each had one customer in mind, so each could pick one set of physical-layer parameters and freeze them — which is exactly what LTE did when it fixed its subcarrier spacing at 15 kHz for every band, every cell size and every speed (M10-L2). 5G NR was specified against three customers at once, and their requirements contradict each other. The interesting content of this lesson is not the list of new features; it is the way that contradiction forced the air interface to become adjustable rather than merely faster.
The three services have names and, unusually for marketing terms, precise numeric definitions. They come from the IMT-2020 requirements (ITU-R M.2410) and 3GPP’s own scenario study, and they are targets a candidate technology had to demonstrate, not measurements of a deployed network — a distinction worth holding on to for the rest of the lesson:
Put each of those in units you can picture. A connection density of 10⁶ per km² is 10⁶ devices divided by 10⁶ square metres, which is one device per square metre, everywhere, across a whole square kilometre. Reliability of 99.999% is one failure in 10⁵ packets, so a machine sending one packet per millisecond — 1000 per second — fails once every 10⁵/1000 = 100 seconds. And a 1 ms latency budget is brutal next to LTE, whose scheduling interval is one 1 ms subframe and whose HARQ retransmission round trip is 8 ms (M6-L4 introduced HARQ as delay bought to save a block): a single LTE retransmission overruns the entire URLLC budget eight times over.
| Corner | Headline requirement | What it demands of the radio | What it is happy to give up |
|---|---|---|---|
| eMBB | 20 Gbit/s peak, 100 Mbit/s experienced | Enormous bandwidth, high-order modulation (256-QAM, M5-L4), many spatial layers (M10-L3) | Latency — a video buffer hides 100 ms without complaint |
| URLLC | 1 ms, 99.999% | Very short transmission intervals, conservative modulation, spare capacity held in reserve | Throughput and spectral efficiency — the packets are tiny |
| mMTC | 10⁶/km², multi-year battery | Narrow bandwidth, deep coverage, a device that sleeps almost all the time | Latency and throughput — a meter reading may wait hours |
Read the last two columns together and the conflict is plain. eMBB wants the widest possible channel and the densest constellation; URLLC wants the shortest possible transmission and margin left unused; mMTC wants the narrowest channel and a radio that is switched off. There is no single set of parameters that is simultaneously best at all three, and 3GPP did not pretend otherwise. Instead NR was specified as a family of configurations sharing one framework, with the subcarrier spacing as the main dial. Flexible numerology is not a feature bullet; it is the direct structural answer to having three customers.
LTE has exactly one numerology, and M10-L2 worked it through: 15 kHz spacing, so a useful symbol lasts Tu = 1/15000 = 66.67 µs, a normal cyclic prefix of 4.69 µs brings the transmitted symbol to 71.36 µs, and the prefix costs 4.69/71.36 = 6.6% overhead. NR keeps every structural decision in that sentence and makes the spacing itself a power-of-two multiple of 15 kHz. Index the choices by k = 0, 1, 2, 3 and the whole numerology falls out of one line:
| Δf | Tu = 1/Δf | Normal CP | Symbol | Slot (14 symbols) | Slots per ms | RB = 12Δf | CP reach |
|---|---|---|---|---|---|---|---|
| 15 kHz | 66.67 µs | 4.69 µs | 71.36 µs | 1 ms | 1 | 180 kHz | 1407 m |
| 30 kHz | 33.33 µs | 2.34 µs | 35.67 µs | 0.5 ms | 2 | 360 kHz | 703 m |
| 60 kHz | 16.67 µs | 1.17 µs | 17.84 µs | 0.25 ms | 4 | 720 kHz | 352 m |
| 120 kHz | 8.33 µs | 0.59 µs | 8.92 µs | 0.125 ms | 8 | 1.44 MHz | 176 m |
Two entries in that table deserve checking by hand. The resource block is always 12 subcarriers — M10-L2 established that and NR did not change it — so its width scales with the spacing: 12 × 120 kHz = 1440 kHz, which is a resource block 1.44 MHz wide, eight times LTE’s 180 kHz. And the slot column is the latency column: the smallest unit a scheduler normally hands out is a slot, so moving from 15 kHz to 120 kHz shortens the scheduling grain from 1 ms to 0.125 ms, an eightfold improvement bought with nothing but a change of variables.
The cyclic prefix is defined as a fraction of the symbol, so it halves whenever the spacing doubles. That has one very convenient consequence and one awkward one. The convenient one is that the overhead is scale-invariant: M10-L2 computed it at 30 kHz spacing, where Tu = 33.33 µs and the prefix is 4.69/2 = 2.34 µs, giving 2.34/35.67 = 6.6% — identical to LTE’s figure, because numerator and denominator scaled together:
The awkward consequence is that the prefix’s duration is what fences off multipath, and duration is exactly what shrinks. At 15 kHz, 4.69 µs of prefix absorbs echoes arriving up to 4.69 µs late, which at 300 m per µs (M10-L4’s convenient constant) is 1407 m of excess path. At 120 kHz it is 0.59 µs and 176 m. So the shortest slots are only usable where the channel’s delay spread is small — small cells, indoors, millimetre wave — and in a large rural cell 120 kHz would reintroduce exactly the inter-symbol interference the prefix exists to prevent. That is the trade in one sentence: shorter slots buy latency and cost multipath robustness.
There is a second reason the spacing has to grow with frequency, and it is the sharper of the two. M8-L3 gave the Doppler shift as fd = v/λ, and inter-carrier interference depends on fd as a fraction of the subcarrier spacing. Take a vehicle at 100 km/h = 27.78 m/s. At 3.5 GHz, λ = 3×10⁸/3.5×10⁹ = 85.7 mm, so fd = 27.78/0.0857 = 324 Hz, which is 324/15 000 = 2.2% of a 15 kHz spacing — comfortable. At 28 GHz, λ = 3×10⁸/28×10⁹ = 10.7 mm, so fd = 27.78/0.0107 = 2593 Hz, which is 17.3% of 15 kHz — unusable — but 2593/120 000 = 2.2% of a 120 kHz spacing. The two figures are identical to three digits, and not by luck: the frequency ratio 28/3.5 = 8 and the spacing ratio 120/15 = 8 are the same number. Scaling the spacing with the carrier frequency keeps the Doppler fraction constant, which is the whole design rule in one line.
The coherence time follows too. M10-L4 needed Tc ≪ the reciprocity interval, and M8-L3’s Tc ≈ 0.423/fd gives 0.423/2593 = 163 µs at 28 GHz and 100 km/h. A 120 kHz slot is 125 µs, so a channel estimate made at the start of a slot is still describing the same channel at the end of it — barely. At 15 kHz the slot would be 1 ms, six times the coherence time, and the estimate would be fiction. Fast numerologies are not a luxury at millimetre wave; they are the only numerologies that work there.
One more latency lever, because the slot table alone does not reach 1 ms end to end. NR allows a transmission to occupy 2, 4 or 7 symbols instead of a whole slot, starting at any symbol boundary rather than waiting for the next slot to begin — officially type-B scheduling, universally called a mini-slot. At 120 kHz a two-symbol mini-slot lasts 2 × 8.92 = 17.8 µs, and even at 30 kHz it is 2 × 35.67 = 71.3 µs. Combined with the shorter slot, this is how the radio part of a 1 ms budget is met: not by transmitting faster, but by being allowed to start sooner and stop earlier.
One honest gap in the table. The specification also defines a 240 kHz spacing, and 480 and 960 kHz were added later for the bands above 52.6 GHz — but 240 kHz is used only for the synchronisation and broadcast signals a device hunts for before it is connected, never for user data. The four rows above are the four numerologies that actually carry traffic in commercial networks, which is why this lesson works with 15, 30, 60 and 120 kHz and mentions the rest only here.
NR splits its spectrum into two frequency ranges, and almost every confused claim about 5G comes from quoting a figure from one range while describing the coverage of the other. FR1 spans 410 MHz to 7125 MHz — colloquially “sub-6” — and carries the network people actually use. FR2 spans 24.25 to 52.6 GHz, the millimetre-wave range, and carries the headlines. M10-L4 already established that FR2 is universally TDD, and that mid-band FR1 is TDD almost everywhere too, so both ranges below get the reciprocity that massive MIMO wants.
| Property | Sub-6 GHz (FR1 mid-band) | Millimetre wave (FR2) |
|---|---|---|
| Typical bands | n77 3300–4200 MHz, n78 3300–3800 MHz, n79 4400–5000 MHz | n257 26.5–29.5 GHz, n258 24.25–27.5, n260 37–40, n261 27.5–28.35 GHz |
| Channel bandwidth | Up to 100 MHz per carrier | Up to 400 MHz per carrier, several aggregated |
| Usual numerology | 30 kHz — 0.5 ms slot, 703 m of prefix reach | 120 kHz — 0.125 ms slot, 176 m of prefix reach |
| Propagation | Diffracts round buildings, penetrates walls with loss (M8-L2) | Barely diffracts; a human body costs a reported 20–40 dB (M8-L2) |
| Cell radius | Hundreds of metres to a few kilometres | Tens to a few hundred metres, line of sight strongly preferred |
| What it is for | Coverage — the layer that carries the traffic | Peak rate and density — stadiums, venues, fixed access |
Now price the difference with Shannon (M6-L1) rather than with adjectives. Give the mid-band cell a favourable 20 dB SNR, which is a linear ratio of 100, and the millimetre-wave cell 10 dB, a ratio of 10 — a fair handicap, because FR2 runs at far higher path loss and lives closer to its noise floor (M9-L1). Both calculations are one line:
The ratio is the lesson: 1384/666 = 2.08, so millimetre wave delivers roughly twice the capacity while carrying four times the bandwidth at ten decibels worse SNR. That is Shannon’s asymmetry made concrete — capacity is linear in bandwidth and logarithmic in SNR, so buying spectrum is a far better deal than buying signal-to-noise ratio, whenever spectrum is available. Turn the comparison around to see how good a deal it is: to get 1.38 Gbit/s out of 100 MHz you would need log₂(1 + SNR) = 1.384×10⁹/10⁸ = 13.84 bit/s/Hz, hence SNR = 213.84 − 1 = 14 650, which is 10 log₁₀(14 650) = 41.7 dB. No mobile link runs at 41.7 dB; even 256-QAM, the densest constellation NR supports (M5-L4), asks for something in the region of 30 dB with coding. The bandwidth is not a shortcut to the same answer — it is the only route to that answer.
Everything so far says millimetre wave should be hopeless, and M8-L1 is where the objection is usually raised: free-space path loss rises as 20 log₁₀f, so moving from 3.5 GHz to 28 GHz costs 20 log₁₀(8) = 18.1 dB at the same distance. Check it the long way at 100 m with the standard formula 32.44 + 20 log₁₀fMHz + 20 log₁₀dkm: at 3.5 GHz that is 32.44 + 70.88 − 20 = 83.3 dB, and at 28 GHz it is 32.44 + 88.94 − 20 = 101.4 dB. The difference is 18.1 dB ✓.
But M8-L1 also supplied the correction that makes millimetre wave viable, and it is the single most important idea in this section: that penalty is an antenna effect, not an absorption effect. The λ in the Friis equation entered through the receiving antenna’s effective aperture, not through the space in between. Hold the antenna gain fixed and higher frequencies lose; hold the physical aperture fixed and higher frequencies win, because aperture gain rises as f². And an array of half-wavelength-spaced elements is precisely a device for holding physical aperture fixed while the wavelength shrinks — you simply fit more elements in:
Make the panel concrete. At 28 GHz, λ = 10.7 mm, so half-wavelength element spacing is 5.36 mm. An 8 × 8 grid of 64 elements spans 7 × 5.36 = 37.5 mm between the outer element centres, so with a little edge the panel is roughly 4.3 cm square, about 18 cm² — a panel that fits behind a lamp-post shroud, and small versions of which fit in a handset. The same 64 elements at 3.5 GHz would need 42.9 mm spacing and span about 34 cm square, some 1180 cm², which is 64 times the area. That factor of 64 is the whole reason massive MIMO (M10-L3) is a millimetre-wave technology first and a mid-band technology second.
Be honest about what the array does not fix. Array gain recovers the free-space term. It does nothing at all about M8-L2’s two harder facts: at λ ≈ 1 cm the wave barely diffracts round an obstacle, and a human body in the path costs 20 to 40 dB, which is a factor of 100 to 10 000 in power. No amount of gain sees through a person. What the array can do is steer — abandon the blocked direction and find a reflected path off a wall — which is why FR2 deployments care about beam management, fast beam switching and having more than one line of sight available. Blockage is the reason millimetre wave is a capacity layer under a sub-6 coverage layer rather than a replacement for it.
M11-L2 described LTE’s architecture as an eNodeB talking to an evolved packet core. NR renames both and changes their internal structure, and the changes are not cosmetic — two of them are what actually deliver the URLLC number, because the radio is only one term in a latency budget.
The 5G base station is the gNodeB, or gNB, and unlike an eNodeB it is not necessarily one box. It is commonly divided into a radio unit at the antenna, a distributed unit nearby handling the time-critical scheduling and physical-layer work, and a central unit that may serve many distributed units from a single site and handles the less time-critical protocol layers. The split matters for a practical reason: the scheduler must answer within a slot — 125 µs at 120 kHz — so it cannot sit at the far end of a long fibre, while the functions that can tolerate milliseconds are cheaper to centralise. The division of labour is chosen by which functions can survive the delay of the link between them.
The 5G core replaces LTE’s fixed set of boxes with a service-based architecture: named network functions that discover one another and communicate over ordinary web protocols rather than over bespoke telecom interfaces. Three of them carry most of the traffic-bearing work, and their separation is the point:
That last clause is the whole latency story. LTE began to separate control from user plane late in its life; NR was designed that way from the start, and because the UPF is an independent function it can be installed at the edge of the network instead of in a national data centre. Price the difference with the same kind of arithmetic M10-L4 used for guard periods, remembering that light in fibre travels at c/n with n ≈ 1.47, so about 2×10⁸ m/s — 200 metres per microsecond rather than the 300 of free space:
So the honest account of how NR reaches 1 ms is a three-part one: shorter slots and mini-slots for the radio, HARQ turned around faster than LTE’s 8 ms, and a user plane physically close to the device. Two of those three are architecture, not radio. This is the single most common omission in explanations of 5G latency, and it is why an operator can deploy every NR feature in this lesson and still not offer URLLC.
Once the core is a set of services rather than a set of boxes, you can instantiate more than one of them. Network slicing is exactly that: several logically separate end-to-end networks running over one physical infrastructure, each with its own selected functions, its own policies and its own service-level agreement. A device signals which slice it wants when it connects, using a slice identifier whose standardised service types are, unsurprisingly, eMBB, URLLC and machine-type communication — the three corners this lesson opened with, promoted into first-class objects in the network.
Be honest about what a slice is and is not. It is genuinely useful: a factory’s control traffic can be given its own session management, its own edge user plane and its own admission policy, so that a stadium full of video streams on the same physical cell cannot starve it. But slicing is an orchestration and commercial capability far more than a radio one. There is still one set of resource blocks in the cell, one power amplifier and one scheduler; a slice does not create spectrum. What it creates is an enforceable claim on the spectrum that exists, plus separate policy, separate charging and separate assurance. Treat “we deployed slicing” as a statement about the core and the operations system, and ask separately what the scheduler was told to do with it.
This is the most misunderstood distinction in 5G, and it explains a specific consumer experience. In non-standalone operation — the way essentially every network launched — a device connects to an LTE base station and a gNB at the same time. The LTE side is the master: it owns the control plane, carries the signalling, manages mobility, and connects to the LTE core. The gNB is a secondary node bolted on to add user-plane throughput. In standalone operation the gNB stands alone with the 5G core, and LTE is not involved at all.
Say the consequence plainly, because marketing did not. A non-standalone device showing a 5G indicator is genuinely using NR carriers for data — the extra throughput is real — while all of its signalling is LTE and its packets traverse the LTE core. That means no 5G core, therefore no service-based architecture, therefore no network slicing; no 5G-native latency guarantees, because the control plane and core were never designed for them; and no voice over NR, so calls fall back to the LTE voice service M11-L2 described. Everything in this lesson that is not purely a radio feature requires standalone. That is why the industry’s slow, expensive migration from non-standalone to standalone matters more to what 5G can actually do than any peak-rate number.
| Aspect | Non-standalone (NSA) | Standalone (SA) |
|---|---|---|
| Control plane | LTE eNodeB, LTE signalling | gNodeB, NR signalling |
| Core network | LTE evolved packet core | 5G core, service-based |
| What you gain | Throughput from the NR carrier, on day one, with existing sites | Slicing, edge user plane, low-latency targets, voice over NR |
| Device indicator | Shows 5G while signalling over LTE | Shows 5G and means it end to end |
M10-L3 made this point about MIMO and it needs making again, harder, here. Work out what the 20 Gbit/s headline actually requires. Take a spectral efficiency of about 6 bit/s/Hz per spatial layer — 256-QAM’s 8 bits per symbol (M5-L4), less the code rate, less the 6.6% prefix, less pilots and control overhead. Then 20×10⁹ / (6 × 4 layers) = 833 MHz of bandwidth, so: two aggregated 400 MHz millimetre-wave carriers, four spatial layers, the densest modulation NR has, and a code rate near one — all at once, and therefore all in line of sight. Multiply it back to check: 800×10⁶ × 4 × 6 = 19.2×10⁹ ≈ 20 Gbit/s ✓.
Every one of those conditions is individually achievable. The problem is the word simultaneously. Line of sight at 28 GHz means no person, vehicle or leaf between you and the panel; a code rate near one means an SNR that only exists close in; four clean spatial layers need the rich scattering that a clear line of sight tends not to provide (M10-L3’s own caveat); and 800 MHz of aggregated FR2 spectrum is deployed in a small number of places. The joint probability of all of them holding is what turns 20 Gbit/s into a demonstration figure. Meanwhile the mid-band cell that serves a whole neighbourhood has one 100 MHz channel, capped by Shannon at 666 Mbit/s at 20 dB SNR, shared among everyone attached to it — and that, not the headline, is the number that describes 5G as most people meet it.
Notice which corner this lesson has not delivered. eMBB is answered by bandwidth, numerology and massive MIMO; URLLC is answered by short slots, mini-slots and an edge user plane. mMTC is answered mostly by something that is not NR at all. The technologies that actually connected the meters and sensors — narrowband IoT and LTE-M — are LTE-derived, narrowband, and were standardised before 5G launched; they are carried forward into 5G networks rather than replaced by an NR design. M11-L4 takes them up alongside Bluetooth, Zigbee and LoRa, and comparing their power budgets against everything in this lesson is the clearest illustration in the course that “newer” and “right for the job” are different questions. Beyond that, Module 12 assembles the whole course into one end-to-end system and looks at what is being proposed for 6G — where the same three-corner argument reappears, with different corners.