Cellular: From 2G to 4G LTE

Module 10 handed you a toolkit and no history. It showed how spectrum can be divided by frequency, by time, by code, by subcarrier and by space, and it left the impression that these are five alternatives sitting on a shelf. They are not. They are five answers that arrived in sequence, each one adopted because the previous answer had run into a wall that the engineers of the day could name precisely. Every cellular generation is a different answer to a question this course has already posed, and this lesson is where the figures you computed in Module 10 stop being exercises and become history.

Five Generations, Five Different Questions

Before the detail, here is the shape of the argument. 1G asked how to give a car a telephone at all, and answered with analog FM in 30 kHz slices. 2G asked how to fit ten times as many users into the same spectrum, and answered by digitising the speech so that eight users could share one carrier in turn. 2.5G asked how to carry data over a network built for calls, and answered with a packet overlay bolted onto the side. 3G asked how to make data a first-class service, and answered with wideband codes. 4G asked how to make the whole network an internet connection that happens to carry voice, and answered by deleting the telephone exchange entirely.

Notice that each answer cost something. Nothing in this sequence is free, and the honest way to read a generation is as a trade the industry chose to accept. Digitising speech cost audio quality at first — early GSM sounded worse than analog AMPS on a strong signal, and users said so. Packet data over a circuit network cost latency measured in hundreds of milliseconds. Wideband codes cost a strict power-control loop that had to run a thousand times a second or the cell collapsed. All-IP cost the telephone network its most reliable service, which is why voice had to be rebuilt from scratch on top of the internet protocol as VoLTE.

1G: An Analog Radio With a Dial Tone

AMPS, deployed commercially from 1983, is the simplest system in this lesson, and M10-L1 already priced it. Each user got a 30 kHz channel carrying frequency-modulated voice — the FM of M4-L1, with a peak deviation of 12 kHz and Carson-rule bandwidth landing inside 30 kHz. Multiple access was pure FDMA: one user, one frequency, for the duration of the call. In the 25 MHz allocated per direction that gives

• 25 MHz ÷ 30 kHz = 833.33, so 833 channels — the figure M10-L1 computed, and about 21 of them were control channels rather than voice
• Switching: circuit, in the most literal sense — a physical path was reserved from the handset to a wire in the telephone exchange and held until someone hung up
• Spectral efficiency for voice: 30 kHz per conversation, and no way to do better, because an analog channel cannot be compressed or statistically multiplexed
• Data rate: none. AMPS carried no digital service at all until the CDPD overlay of 1993 squeezed 19.2 kbit/s into idle 30 kHz channels — 19.2/30 = 0.640 bit/s/Hz

What it solved: mobility, for the first time, at scale. What it cost: everything else. No encryption, so a scanner could listen to any call. No authentication, so cloning a handset’s identity was trivial and became an industry. And 833 channels is not many — a busy city ran out of them, and the only remedy available to an analog FDMA system is to build more, smaller cells and reuse frequencies harder, which M8-L1’s path loss and M9-L3’s co-channel interference put a firm ceiling on.

2G: Digital, and Still a Circuit

GSM, first service 1991, answered the capacity question by moving the voice into bits. Once speech is a bitstream you can compress it, protect it with the coding of M6-L1, encrypt it, and — crucially — interleave several users onto one carrier in time. M10-L1 gave the frame arithmetic and it is worth restating, because every later generation is measured against it: a 200 kHz carrier runs at 270.833 kbit/s gross, divided into 8 time slots of 577 µs each carrying 156.25 bits.

So GSM’s per-carrier efficiency is 270.833/200 = 1.354 bit/s/Hz, exactly as M10-L1 found, and its voice capacity is 200 kHz ÷ 8 = 25 kHz per conversation against AMPS’s 30 kHz. Read that comparison honestly: it is a gain of only 30/25 = 1.2×. The famous three-to-six-fold capacity advantage of GSM over AMPS did not come from the slot structure at all. It came from tighter frequency reuse — digital signals tolerate more co-channel interference (M9-L3) than analog FM does, so cells can reuse the same frequency closer together — and later from the half-rate codec, which put two users in one slot pair.

And the switching stayed circuit. A GSM traffic channel is a standing reservation: your slot in every frame, 217 frames per second, whether you are speaking or silent. For a conversation that is defensible. For data it is a disaster, and the first attempt proved it: Circuit Switched Data gave 9.6 kbit/s by handing a traffic channel to a modem, billed by the minute, dialling up like a landline, and holding the whole channel open while you read the page it had fetched.

The packet overlay: GPRS, then EDGE

GPRS (2000) is the moment packet switching enters cellular, and it is instructive precisely because it is a bolt-on. The radio slots are the same 577 µs GSM slots; what changed is that they are no longer reserved. A handset is given slots when it has something to send, and gives them back when it does not, so many users share the same slots statistically. Four coding schemes trade protection for rate: CS-1 at 9.05, CS-2 at 13.4, CS-3 at 15.6 and CS-4 at 21.4 kbit/s per slot, the last with no error correction at all and therefore usable only close to the base station — which is the adaptive modulation and coding idea of M6-L4 in its crudest form, chosen per connection rather than per millisecond.

EDGE (2003) went further and changed the modulation. GSM uses GMSK, one bit per symbol; EDGE adds 8PSK at three bits per symbol — the phase-shift keying of M5-L3, and the same order-versus-robustness trade M5-L4 laid out. Nine modulation and coding schemes replace the four, MCS-1 to MCS-9, and the top one carries two 592-bit blocks per 20 ms radio block:

• MCS-9 per slot: 2 × 592 bits ÷ 20 ms = 1184/0.020 = 59.2 kbit/s
• The advertised EDGE peak, 236.8 kbit/s, is a four-slot handset: 4 × 59.2 = 236.8 kbit/s ✓. Most handsets were class 10 or class 12, four downlink slots at most
• All eight slots would give 8 × 59.2 = 473.6 kbit/s, which is the specification’s ceiling and was rarely a product
• GPRS for comparison: CS-4 on eight slots is 8 × 21.4 = 171.2 kbit/s theoretical, while a real four-slot handset on CS-2 got 4 × 13.4 = 53.6 kbit/s — which is why GPRS in practice is quoted at 40–60 kbit/s
• Latency stayed poor: 600–700 ms round trip on GPRS, 300–400 ms on EDGE, because the packet path was threaded through a network whose timing was designed around a 4.615 ms voice frame and a hierarchy of controllers

3G: Codes, and Data as a First-Class Citizen

UMTS, from 2001, is M10-L1’s CDMA built at scale, and its distinctive number is the chip rate. Every user’s bits are multiplied by a spreading sequence running at 3.84 Mcps inside a 5 MHz carrier, so all users occupy the whole carrier all the time and are separated by their codes rather than by frequency or slot. M10-L1 computed the processing gain for a 12.2 kbit/s speech service, and it is the number that makes the scheme work at all.

That trade is 3G’s defining property, and it produced two features nothing before it had. The first is soft capacity: because users are separated by code and not by an allocated slot, there is no hard limit of 8 or 833. One more user simply raises the noise floor for everyone, so the cell degrades gradually instead of rejecting the call — and the operator can trade quality for a few more connections at the busiest minute. The second is the cost of that: fast power control. If one nearby handset shouts, its code’s cross-correlation buries every distant user, so the network commands each handset’s power up or down 1500 times a second. Get that loop wrong and the cell does not degrade — it falls over.

R99 UMTS delivered 384 kbit/s, and its per-user spectral efficiency is unflattering: 384/5000 = 0.0768 bit/s/Hz, an order of magnitude below EDGE’s 2.368. That is not a mistake in the design; it is what spreading is. The 5 MHz buys immunity to the multipath of M8-L3 and room for dozens of simultaneous users, not a big number for one of them. HSPA (from 2005–2007) fixed the headline figure by adding 16-QAM, a 2 ms transport interval, fifteen parallel codes and hybrid ARQ — the HARQ of M6-L4, appearing here for the first time in a commercial system. Category 10 HSDPA carries 27 952 bits per 2 ms interval, so 27 952/0.002 = 13.98 Mbit/s, the figure marketed as 14.4 Mbit/s, and 13.98/5 = 2.80 bit/s/Hz.

4G LTE: Subcarriers, and No Circuit at All

LTE, standardised in Release 8 and in service from 2009, made three changes at once, and each one is a lesson you have already had. Multiple access became OFDMA (M10-L2), so users are given rectangles of subcarriers and symbols rather than codes. Antennas became plural: MIMO (M10-L3) is not optional in LTE, it is assumed by the peak-rate definitions. And the switching became all-IP, with no circuit-switched domain anywhere in the specification. M10-L2 gave the numerology and it is the frame every LTE figure in this lesson rests on:

• Subcarrier spacing 15 kHz, so the useful symbol is 1/15 000 = 66.67 µs
• Cyclic prefix 4.69 µs, so the symbol period is 66.67 + 4.69 = 71.36 µs and the overhead is 4.69/71.36 = 6.6%
• Resource block = 12 subcarriers × 15 kHz = 180 kHz, by 7 symbols — the smallest thing the scheduler can hand out
• A 20 MHz carrier holds 100 resource blocks, so 100 × 180 kHz = 18 MHz occupied, leaving 2 MHz (10%) of guard at the edges
• 14 symbols per 1 ms subframe: 14 × 71.36 = 999 µs, and the missing microsecond is the longer first cyclic prefix of each slot (5.21 µs rather than 4.69)

The peak rate, checked against M6-L3

Build the peak downlink rate from the numerology rather than quoting it. A 20 MHz carrier has 100 × 12 = 1200 subcarriers; each carries 14 symbols per millisecond; 64-QAM (M5-L4) puts 6 bits on each symbol; and 2×2 MIMO (M10-L3) sends two spatial streams at once.

That 7.5 is worth stopping on, because M6-L3 predicted it from a completely different direction. That lesson estimated a 2×2 LTE link at roughly 7.5 bit/s/Hz from Shannon capacity and a practical implementation margin, with no reference to resource blocks or category tables. Here the same number falls out of 1200 subcarriers, 14 symbols, 6 bits and 2 streams minus a quarter of overhead. Two independent routes to 150/20 = 7.5 is the kind of agreement that tells you both routes are sound — and it is also the reason the Cat-4 handset became the industry’s reference point rather than an arbitrary choice.

Circuit Versus Packet, and Why All-IP Won

A circuit is a reservation. GSM gave you a slot in every frame for the length of the call, and the cost of a reservation is that it is paid whether you use it or not. For conversation that is nearly efficient, because speech occupies the channel roughly half the time in each direction and the reservation guarantees the 20 ms cadence a codec needs. For anything else it is ruinous. Loading a web page is a burst of a few hundred kilobits followed by seconds of nothing while you read; a dedicated 13 kbit/s channel spends that reading time transmitting silence at full cost.

Put the two side by side on the same slot. Circuit switching gives one user a slot permanently: utilisation equals that user’s duty cycle, so a reader who is active 5% of the time uses 5% of a reserved channel. Packet switching gives the slot to whoever has data this instant, so twenty such users can share one slot and each still gets the full 21.4 kbit/s during their burst. The multiplexing gain is exactly the ratio of peak to average demand, and for interactive data that ratio is enormous — which is why every generation from GPRS onward moved further towards packets and 4G went all the way.

• Circuit: guaranteed rate and delay, zero multiplexing gain, billed by time. Correct for a 1980s voice network
• Packet: no guarantee without an explicit priority mechanism, large multiplexing gain, billed by volume. Correct for anything bursty
• The hybrid (2.5G): packets carried over a circuit-era network, which inherits the latency of the hierarchy it was bolted onto — 600–700 ms on GPRS
• All-IP (4G): there is no circuit domain to fall back on, so the priority mechanism is not optional — it becomes the QCI system that VoLTE depends on below

The Flat Architecture

Here is the change a learner is most likely to miss, because it is not a radio change at all. UMTS put a Radio Network Controller between the base station and the core: the NodeB was a comparatively dumb radio, and the RNC above it held the scheduler, the handover decisions and the radio-resource control state for a hundred cells at once. LTE deleted that box. The eNodeB holds its own scheduler, its own RRC state machine and its own handover logic, and talks to its neighbours directly over an interface called X2 rather than through a parent.

Two structural facts fall out of that table. First, MME plus S-GW/P-GW is a separation of the control plane from the user plane: signalling and data travel through different boxes, so an operator can scale the two independently and a signalling storm cannot congest anybody’s download. Second, and say it plainly: removing a hierarchy layer is what cut the latency. A 3G round trip of roughly 150 ms became 30–50 ms in LTE, a factor of about 150/40 = 3.75, and the radio contributed only part of that. Every packet in UMTS crossed NodeB → RNC → SGSN → GGSN; in LTE it crosses eNodeB → S-GW → P-GW. One fewer store-and-forward hop, each with its own queue and its own processing, and the queue you removed was the busiest one.

The 1 ms TTI and the scheduler

The flat architecture is what makes the LTE scheduler possible, and the scheduler is what makes M6-L4 real. LTE’s transmission time interval is 1 ms — one subframe — and within every single one of them the eNodeB decides, for every active user, which resource blocks they get, at which modulation order, with which code rate. That is 1000 decisions per second per user, and it cannot be done from a controller a hundred kilometres away because the information it needs is stale by the time it arrives.

The information is the channel quality indicator. Each handset measures the downlink and reports a CQI from 1 to 15, and the eNodeB maps that index onto the ladder M6-L4 built: QPSK at the bottom, then 16-QAM, then 64-QAM, with the code rate climbing inside each step. M6-L4 measured the span of that ladder at a factor of about 36 in throughput over the same bandwidth, and HARQ sits underneath it as the cheap correction when the eNodeB guesses one step too high. So the answer to “why is LTE fast?” is not only OFDMA and MIMO. It is that a box at the edge of the network re-optimises the link a thousand times a second, which no 3G RNC could have attempted.

VoLTE: Voice as Just Another Packet

Now the consequence that surprises people, and it is a genuine surprise rather than a piece of trivia. LTE has no circuit-switched voice service whatsoever. There is no traffic channel, no MSC in the path, nothing in Release 8 that carries a telephone call as a call. The system that replaced the mobile telephone network cannot, by itself, make a telephone call — and for the first several years of commercial LTE that was literally true: an LTE handset with a strong data connection dropped to 2G or 3G whenever it rang.

That fallback is CSFB, circuit-switched fallback: on an incoming call the MME pages the handset, tells it to move to the legacy network, the call is set up there, and the LTE data session is suspended or handed over. It works, and it costs one to two seconds of extra call setup plus a legacy network the operator wanted to switch off. VoLTE is the real answer: the speech is an ordinary IP flow, RTP over UDP over IP, carried on a dedicated bearer whose QCI is 1 — guaranteed bit rate, packet delay budget 100 ms, tolerated packet loss 10⁻² — with IMS signalling on QCI 5 and its far stricter 10⁻⁶ loss target. The QCI is the whole trick: without it, a voice packet queues behind a video download and the call is unusable.

And the packetisation has a cost that is easy to compute and startling to see. AMR-WB at 12.65 kbit/s produces 12 650 × 0.020 = 253 bits = 31.6 bytes every 20 ms, and the headers on top are RTP 12 + UDP 8 + IPv4 20 = 40 bytes.

• No circuit domain in LTE at all — a genuine consequence of all-IP, not an oversight
• CSFB hands the call to 2G/3G, costs 1–2 s of setup, and requires keeping the old network alive
• VoLTE carries speech as RTP/UDP/IP on a QCI-1 bearer: guaranteed rate, 100 ms delay budget, 10⁻² loss
• ROHC is mandatory in practice: 40 bytes of header on 31.6 bytes of speech is 126% overhead, and 3 bytes is 9.5%

The Arc, in One Table

Every figure below is the peak downlink rate one carrier of the stated width could deliver to a single user, divided by that width. State the definition, because it matters: a per-carrier efficiency would flatter GSM (270.833/200 = 1.354 bit/s/Hz across all eight users) and a per-user figure flatters nobody in the early generations.

Take the endpoints of the digital story: GSM’s 13 kbit/s speech codec in 1991 to LTE Cat-4’s 150 Mbit/s around 2012 is 150 000/13 = 11 538× in 21 years — four orders of magnitude. Compounded, 11 538^1/21 = 1.56, so 56% per year, a doubling every ln 2 / ln 1.56 = 1.6 years. That is a semiconductor-industry growth rate sustained by a radio industry for two decades, and the next section says where it actually came from.

Where the growth actually came from

Compare the first and last rows of the table on both measures. The peak rate went from 19.2 kbit/s to 150 Mbit/s, a factor of 150 000/19.2 = 7813×. The spectral efficiency went from 0.640 to 7.500 bit/s/Hz, a factor of only 7.5/0.640 = 11.7×. Those two ratios must differ by exactly the bandwidth ratio, and they do: