Cellular modem engineering

2G to 6G in one deck: what a modem is, how each generation reshaped it, and what a 6G modem has to do.

Status as of 21 October 2026. Arrow keys move between slides; G opens the slide grid.

Route through the deck

Foundations first, then each generation, then everything that makes 6G different, then Apple and the interview.

Tags: Sourced from a cited page, Analysis my synthesis, Reported press reporting not confirmed by the company.

Forty years of cellular telephony in one row

Each generation changed the tower, the phone and what people did with it. Illustrations are generic silhouettes, not product renders.

Years are first commercial launches. The 6G column is an expectation, not a finished standard. Analysis

A modem turns bits into radio and radio into bits

Modem means modulator plus demodulator. A cellular modem also runs the whole conversation with the network, so it is a radio, a signal processor and a protocol engine in one package.

Application processorApps, OS, IP stack. Hands packets to the modem over PCIe, USB or shared memory.

Layer 3RRC and NAS: connection control, registration, bearers, mobility decisions.

Layer 2MAC, RLC, PDCP: scheduling grants, HARQ, segmentation, ciphering, header compression.

Baseband PHYModulation, coding, FFT, MIMO detection, channel estimation, synchronization.

RF transceiverADC and DAC, mixers, PLLs, filters. Moves between baseband and carrier frequency.

RF front endPower amplifier, LNA, switches, filters, duplexers, antenna tuners.

AntennasSeveral per band; arrays and beams for millimeter wave.

Digital: software, DSP, fixed-function acceleratorsAnalog and RF: filters, amplifiers, antennas

Also insideSIM or eSIM and key handling, clocks, power management interface, GNSS coexistence, and test and calibration hooks.

The permanent trade-offThroughput against power, silicon area, heat and cost, across dozens of bands and up to five generations at once.

From power-on to a moving data session

Click each stage. This sequence is the same in every generation; only the signals and message names change.

Say it in an interview: the modem is a closed-loop system. It measures the channel, reports it, and the network adapts. Most modem features are one side of a loop.

The same conversation as a sequence of messages. Names are LTE/NR style.

The channel sets the ceiling

Multipath fading, Doppler, noise and interference make the radio channel hostile. Shannon gives the ideal upper bound on what any modem can get out of it.

BandwidthCosts spectrum and RF and baseband power. Doubling it doubles capacity.

SNRCosts link budget and interference control. At high SNR each 6 dB adds about 1 bit per hertz per layer.

Spatial layersMIMO multiplies capacity when the channel has rank. Costs antennas, compute and power.

Bandwidth SNR Spatial layers

ideal upper bound, C = layers × B × log2(1 + SNR). Real modems reach a fraction of it.

One transmitted signal arrives as several copies with different delays (multipath) and a frequency shift that depends on speed (Doppler).

2G: digital voice, shared by time

GSM digitized voice and let eight users take turns on each 200 kHz carrier. Time slots separate users; different carriers separate cells.

Access TDMA + FDMACarrier 200 kHzModulation GMSK, 8PSK in EDGEData 9.6 kbps circuit; GPRS up to ~171 kbps; EDGE up to ~384 kbps

What the modem paysEqualizer (Viterbi / MLSE) against multipath, frequency hopping, timing advance. Small DSP, tiny power budget.

Say it: 2G taught the core timing lesson. Users far from the tower must transmit early (timing advance) so their slot lands in the right place.

Each slot lasts about 577 µs. The highlighted slot belongs to one phone; it transmits a burst once per frame.

3G: everyone talks at once, separated by codes

UMTS spread every user across the full 5 MHz. Unique spreading codes keep them apart, so interference, not time slots, becomes the capacity limit.

Access WCDMACarrier 5 MHz, 3.84 McpsR99 384 kbps typicalHSDPA ~14 Mbps, 16QAM, HARQ, 2 ms TTIHSPA+ 42 Mbps+, 64QAM, MIMO, dual carrier

What the modem paysRake receiver to combine multipath, closed-loop power control at 1500 Hz, soft handover, chip-rate processing, 2G fallback. Multi-mode begins here.

Say it: CDMA trades frequency planning for power control. The near-far problem is why every phone must transmit only as loud as needed.

Receivers recover one user by correlating with that user's code; the others look like noise.

4G: many narrow subcarriers, one flat channel each

LTE splits the band into 15 kHz subcarriers that are orthogonal. Each sees a flat channel, so one complex multiply per subcarrier equalizes it. A cyclic prefix absorbs delay spread.

Downlink OFDMAUplink SC-FDMACarrier up to 20 MHzTTI 1 msMIMO up to 4×4, 256QAM, turbo codeLTE-A carrier aggregation

What the modem paysFFT engines, MIMO detection, turbo decoder throughput, dozens of bands, carrier-aggregation band combinations in the RF, IMS for VoLTE.

Say it: uplink uses SC-FDMA because a lower peak-to-average power ratio lets the phone's amplifier run efficiently. That is battery and coverage.

Each curve is one subcarrier. At every subcarrier's peak, all the others are exactly zero, so they do not interfere.

5G NR: flexible numerology, beams, two spectrum worlds

NR keeps OFDM but makes it configurable. Subcarrier spacing scales, slots shrink for low latency, and millimeter wave adds beamforming as a core modem job.

FR1 up to 100 MHz carriersFR2 24.25–52.6 GHz, up to 400 MHzCoding LDPC data, Polar controlModes NSA (EN-DC) and SAIMT-2020 20 Gbps, ~1 ms

What the modem paysBeam management, wide-bandwidth ADCs and FFT, LDPC throughput, two radios active under EN-DC, mmWave heat, UE power-saving features.

Say it: subcarrier spacing = 15 × 2^μ kHz, slot = 1 / 2^μ ms. Higher μ is for millimeter wave and low latency.

μ	0	1	2	3
Spacing	15 kHz	30 kHz	60 kHz	120 kHz
Slot	1 ms	0.5 ms	0.25 ms	0.125 ms

Each generation bought roughly 10 to 50 times the peak rate

Peak rates are specification targets under ideal conditions. They show direction and ratio, not what a user sees.

Peak data rate, log scale. 2G is EDGE, 3G is HSPA+ (R8), 4G is LTE-Advanced, 5G is the IMT-2020 target, 6G is the top of the IMT-2030 range (50 to 200 Gbps by scenario).

Single-carrier bandwidth, log scale. LTE aggregates carriers beyond 20 MHz; NR FR2 reaches 400 MHz per carrier.

Say it: rate = bandwidth × spectral efficiency × layers. Look at which factor each generation pulled: 3G bandwidth, 4G spectral efficiency and layers, 5G bandwidth again.

The evolution thread, side by side

	2G	3G	4G	5G	6G (direction)
Access	TDMA / FDMA	CDMA	OFDMA, SC-FDMA	OFDMA, DFT-s-OFDM uplink option	CP-OFDM downlink; CP-OFDM and DFT-s-OFDM uplink Sourced
Carrier	200 kHz	5 MHz	≤ 20 MHz	≤ 100 MHz (FR1), ≤ 400 MHz (FR2)	Wider carriers; upper mid-band in focus
Coding	Convolutional	Convolutional, turbo	Turbo	LDPC data, Polar control	5G codes largely reused Sourced
Antennas	1	1–2	≤ 4×4	Massive MIMO, beams	Larger arrays, extreme MIMO under study
Latency unit	4.6 ms frame	2 ms TTI (HSDPA)	1 ms TTI	≤ 1 ms slot, mini-slots	Targets set by IMT-2030 HRLLC
Core	Circuit	Circuit + packet	All-IP EPC	Service-based, cloud-native	New 6G core, AI and sensing aware
Modem pays	Equalizer	Rake, power control	FFT, MIMO, turbo, CA	Beams, LDPC, mmWave heat	AI inference, wider RF, sensing Analysis

What limited each generation

2GVoice only, circuit-switched.

3GInterference between users sharing the band.

4GSpectral efficiency and bandwidth per carrier.

5GPeak rate, latency and flexibility across very different bands.

6GEnergy per bit, coverage at new frequencies, and intelligence in the loop. Analysis

Patterns to internalize

Each generation raises bandwidth and spectral efficiency; the modem pays in compute and power.
Intelligence moves toward the edge: fast scheduling and HARQ in 3G, beam management in 5G, learned PHY functions next.
Backward compatibility is a tax: a 6G phone modem will be five-mode.
New spectrum forces new RF architecture and coexistence work.
In a phone, power and heat limit throughput long before the standard does.

The clock to 6G

3GPP is using two releases: Release 20 for studies, Release 21 for the first normative 6G work. Commercial service is widely expected around 2030.

Dates move. A recent 3GPP plenary readout puts the Release 20 6G study running through June 2027; earlier plans showed a different Release 21 schedule. Check the 3GPP work plan before quoting a date. Sourced The ITU bar is approximate. Analysis

The 6G air interface looks a lot like 5G

The headline from the first agreements: the physical layer is evolutionary. The big changes are in system design, AI, sensing and spectrum.

Settled or agreed Sourced

CP-OFDM in the downlink; CP-OFDM and DFT-s-OFDM in the uplink. DFT-s-OFDM gives better phone coverage.
5G channel codes (LDPC and Polar) largely reused, with possible extensions.
A new radio access technology with standalone operation, not backward compatible with NR (called 6GR in the study).
Avoid defining several numerologies for the same frequency band.
AI/ML must have a non-AI fallback so service holds when AI features are absent.
Sensing is studied as a native air-interface capability: waveform, reference signals, measurement feedback.

Still open or to verify Analysis

Modulation details, reference signal structure and supported bandwidths: the milestone list names them, so read the outcome in the RAN1 study report (TR 38.760-1).
Which AI/ML use cases become normative and how models are managed across vendors.
How sensing coexists with communication in a phone modem.
RAN and core functional split, and how NTN fits natively.
Security framework beyond control-message protection.

Key documents: TR 38.914 (scenarios and requirements), TR 38.760-1 (6G radio, RAN1), TS 22.270 (system requirements, draft), TR 22.870 (use cases), TR 23.801-01 (architecture).

Say it: if asked what changes in the modem, do not start with a new waveform. Start with wider spectrum, AI with fallback, sensing, NTN and energy per bit.

What ITU asks of 6G

Three scenarios extend 5G (violet); three are new (pink). Recommendation ITU-R M.2160. Sourced

IMT-2030 sets 15 capabilities: 9 enhanced from 5G and 6 new (coverage, positioning, sensing, AI, sustainability, interoperability). ITU's working party has agreed a draft of 20 technical performance requirements.

Targets vary by usage scenario; bars show the 6G range against the 5G figure. Latency: IMT-2020 set 1 ms user-plane; the 6G number depends on the scenario and is still in the requirements work.

Where 6G is looking for spectrum

Log frequency axis. WRC-27 agenda item 1.7 studies 4.4–4.8, 7.125–8.4 and 14.8–15.35 GHz; WRC-23 identified upper 6 GHz (6.425–7.125) for IMT in some regions. Sourced

Why the 7 to 15 GHz rangeMore than 2 GHz of possible spectrum with propagation closer to bands below 7 GHz than to millimeter wave, so existing sites can be reused. The costs: weaker uplink coverage and larger arrays.

It is not settledRegulators differ: CEPT and the UK have been skeptical of the case for more spectrum, while the US is pushing 7 and 8 GHz. Outcomes come at WRC-27 in late 2027.

What a higher carrier does to the link

Frequency Speed Distance

Free-space path loss between isotropic antennas at the chosen distance.

What a 6G modem has to handle

Pick an area. Each entry says what the modem must do, why it is hard, and the question an interviewer is likely to ask.

An illustrative 6G modem

Pink blocks are what changes against a 5G modem. This is my synthesis from the study direction, not an Apple or 3GPP design. Analysis

Host

Application processor, OS, IP stack

Layers 2 and 3

RRC, NAS, mobility

MAC, RLC, PDCP, HARQ

AI/ML lifecycle: model selection, monitoring, switch to fallback

Physical layer

FFT / IFFT, DFT spreading

MIMO detection, precoding

LDPC, Polar codecs

Sync, channel estimation

Neural inference path: CSI, beams, channel estimation

Non-AI fallback path (required)

Sensing pipeline: reference signals, measurements

RF

Sub-6 GHz paths, 2G to 5G

mmWave paths

Upper mid-band path, wide ADC / DAC

NTN: Doppler and delay pre-compensation

Design pressure to remember: every pink block draws power and area, and all of them share one thermal budget in a thin phone.

Where Apple's own modem stands

Feb 2025: C1First in-house modem, in iPhone 16e. Apple calls it the most power-efficient modem ever in an iPhone. Sub-6 GHz 5G only. Sourced

Sep 2025: C1XVariant in iPhone Air, later iPhone 17e. Sourced

Sep 2026: C2In iPhone 18 Pro; adds mmWave and is reported to use about 15% less power than C1X. Sourced

Not fully switched overA teardown of the iPhone 18 Pro Max reportedly found a Qualcomm Snapdragon X80, so Apple's modem does not yet cover the whole lineup. Reported

2027 and later: C3, SoC integrationC3 (codename Prometheus) is expected in 2027; reports describe a longer plan to fold the modem into Apple silicon. Reported

What this means for a modem engineer

Apple's pitch is efficiency and integration with its own silicon, so expect questions about energy per bit, thermal limits and power states.
The team is still closing feature gaps with the incumbent (mmWave arrived in C2), so breadth across the standard matters.
6G prep is a long game: first normative specs arrive around 2029, and a phone modem follows later. Show you can reason from the study documents.

Earlier Bloomberg reporting said Apple had posted 6G-related roles, including a 6G reference architecture position. Reported

Apple confirms its own products and claims; roadmap items beyond that come from press and can change.

iPhones through the cellular generations

Generic silhouettes by design era, not product renders. First-iPhone-per-generation dates from Apple's public history; C1 per Apple.

24 questions to rehearse out loud

Click a card to reveal a model answer. Say your own answer first.

Formulas you should be able to derive

Shannon capacity

C = B · log₂(1 + SNR)

MIMO: about min(Nt, Nr) layers × this, when the channel has rank. 100 MHz at 20 dB: 100 × 6.66 = 666 Mbps per layer.

OFDM timing

T_sym = 1 / Δf, Δf = 15 · 2^μ kHz

15 kHz → 66.7 µs useful symbol. Cyclic prefix must exceed delay spread.

Doppler

f_d = v · f_c / c

500 km/h at 7 GHz: 138.9 m/s × 7 GHz / c ≈ 3.24 kHz, about 22% of a 15 kHz subcarrier.

Free-space path loss

FSPL = 20 log₁₀(4π d f / c)

Doubling frequency adds 6 dB between isotropic antennas. Same-area arrays recover it with beamforming gain.

Noise floor and sensitivity

N = −174 dBm/Hz + 10 log₁₀(B) + NF

20 MHz, NF 7 dB: −174 + 73 + 7 = −94 dBm. Sensitivity = N + required SNR.

Subcarriers and sampling

f_s = N_FFT · Δf

LTE 20 MHz: 2048 × 15 kHz = 30.72 MHz. Check your FFT size against the carrier before sizing the ADC.

What to study next

Tick items as you finish them; progress is saved in this browser.

Sources

Standards and bodies

Spectrum and Apple

The 2G to 5G generation figures, the formula sheet and the interview answers are standard engineering knowledge written from memory and not individually sourced. Verify exact figures against 3GPP specifications before quoting them in an interview.