Evolution notes for this page (click to expand)

Wireless Evolution

7 stages, smallest to largest; each one starts from what broke in the previous version.

① 1G: analog cellular

~56 voice ch./site · 1980s

This architecture: ★ 1G swaps one high-power tower for many low-power cells plus frequency reuse: cells far enough apart share a channel group, sites backhaul to the MTSO (later the MSC), which assigns channels, orders handoffs and reaches the PSTN. Voice is analog FM; the control channel is already digital FSK.

What broke in the previous stage

  • ★ First, the premise: the generation before 1G is not another mobile network but the pre-cellular single tower (MTS 1946, IMTS 1964). One several-hundred-watt tower covered a whole city, so while a call held a channel nobody else in the city could use it — capacity equalled the channel count, about a dozen per city. The quoted 1976 New York IMTS snapshot: a few hundred paying subscribers, thousands waiting, about a dozen calls at once. Growth needed spectrum, and the new band, carved out of UHF TV in 1970, sat in a decade of rule-making and litigation; cellular licensing rules landed only in 1981.
  • ★★★ The counter-intuitive part: from 1G on, capacity is no longer set by spectrum alone but mainly by cell count. Cut each transmitter to just cover its own cell and cells D=R√(3N) apart can reuse the same group (AMPS often N=7; a site nominally gets ~56 voice channels). Growth comes from cell splitting: halve the radius, four times the cells. The gain stays below 4x — fewer channels per cell means worse trunking (56 channels carry ~45 Erlang at 2% blocking; four cells of 14 carry clearly less) — and site cost plus the co-channel noise floor put an economic and physical floor under splitting. That reuse factor of 7 is exactly what 2G attacks next.
  • ★★ The misreading worth fixing: calling 1G analog means the voice (traffic) channel uses analog FM — not that everything is analog. Access, paging and channel assignment ride a 10 kbps digital FSK control channel; blank-and-burst is used only for in-call handoff orders and power control. But that digital pipe is fixed-format and tiny, far too small for user services. The real killer is identity: ESN/MIN went out in the clear and the original AMPS spec had no challenge-response, so a captured pair could be burned into a clone (US losses ran about 0.5-0.65 billion USD a year in 1995-1996). On the voice side, demodulating FM gives audio — there is no bit stream to encrypt.

② 2G · Going digital

Theoretical peak ~474 kbit/s · 1990s-2000s

This architecture: ★ 2G turns the air interface into digital bits: handset plus SIM reaches a BTS over Um, the BSC owns radio resources and handover, MSC/VLR carries circuit voice to the PSTN. ★★ From 2.5G, SGSN/GGSN add a packet domain beside it.

What broke in the previous stage

  • Start from what the 1G page actually stated: a call owned one 30 kHz channel each way, the 2:1 compander improved SNR but saved no bandwidth, so no bandwidth could be shared mid-call — about 56 nominal voice channels per AMPS site; handover was measured network-side and ordered by the MTSO; ESN/MIN went out in clear, so copying meant cloning. ★★★ One root cause: the waveform was still analog — no coding layer to compress, no bitstream to compute on. ★ New figure here: FM needed ~17 dB C/I, so reuse stayed at N=7.
  • ★★ The usual miscalculation is that digitization saves bandwidth. GSM full rate squeezes speech to 13 kbit/s and a 200 kHz carrier holds 8 slots, so a call takes 25 kHz — barely under AMPS's 30 kHz, and TACS was already 25 kHz, NMT-900 even 12.5 kHz: zero gain there. The multiplier comes from coding cutting the interference threshold to ~9 dB in spec (~12 dB in planning), which lets reuse tighten from 7/21 to 4/12 or 3/9, and to 1/3 with hopping. ★★★ You save C/I margin, not bits.
  • ★★★ Two cuts to remember. The SIM pries identity out of the device: IMSI and Ki never leave the card, A3/A8 run inside it, the number lives in the HLR — so you swap handsets and keep the number, prepaid needs no contract, and roaming uses the IMSI's MCC/MNC to find the home HLR. IS-95 burned identity into the handset, so card swapping never caught on there. GPRS changed the switching mechanism (20 ms radio blocks, billing by volume); EDGE only changed the rate. ★ Always-on is not holding radio: the context persists, resources come on demand.

③ 3G · Going packet

R99 384k → HSPA 1-3M · 2000s

This architecture: ★★★ 3G changes only the access side: Node B plus RNC and a 5 MHz spread-spectrum air interface; the CS / PS core is inherited from 2G GPRS. What made it usable was HSPA - shared channel, 2 ms TTI, scheduling and HARQ moved down into the Node B.

What broke in the previous stage

  • ★★★ Order first: WCDMA was specified in 1998-1999, before GPRS went commercial, so 3G was not forced out by GPRS - the drivers were IMT-2000 planning, voice capacity hitting its ceiling and new spectrum. HSPA is the generation lived experience forced. ★★★ The 2G page only gave a peak of about 474 kbit/s, an arithmetic figure that hands all 8 slots to one user. The real mechanism is reservation: a 200 kHz carrier is cut into 8 slots, one full-rate call owns a TCH from dial to hang-up (DTX does not release the slot), and slots borrowed for packet data are reclaimed for voice once the cell is busy.
  • ★★★ The easiest thing to get wrong: everything 3G adds sits on the access side - Node B, the RNC (successor to the BSC) and a 5 MHz spread-spectrum air interface. The core is inherited: CS / PS domains and SGSN / GGSN were built by GPRS in 2G, and one core serves both the 2G BSS and the RNC. So credit the break from circuits - holding an IP address is not holding a radio channel - to GPRS; 3G added cheap camped states (CELL_PCH / URA_PCH). ★★ And 3G is not the same as CDMA spreading: by the standard, EDGE counts too.
  • ★★★★ Take this away: HSPA made 3G usable, not R99. R99's 2 Mbit/s needs multicode and eats most of the code tree, so commercial networks rarely granted it; starting at 2.1 GHz, cell-edge users often got less than EDGE. HSPA moved data onto the shared HS-DSCH, cut the TTI to 2 ms and pushed scheduling and HARQ into the Node B: around 2008 the field saw 1-3 Mbit/s and RTT 70-150 ms. ★★★ What saturated first was the number of connection setups - RNC control plane and Iub signalling full with the user plane half empty. That is why LTE removed the RNC.

④ 4G LTE · all-IP

peak ~150 Mbps · 2010s

This architecture: ★★★★ A flat all-IP network: the eNodeB schedules and negotiates handover over X2; the RNC layer is gone. The EPC keeps only MME and S/P-GW; the air interface is OFDMA + MIMO. ★ LTE has no circuit domain — voice rides VoLTE or falls back to 2G/3G.

What broke in the previous stage

  • ★★★★ Premise first: 3G here includes HSPA+, not just the previous page's Rel-99. WCDMA is code-division — one 5 MHz carrier shared by the whole cell, users told apart by codes. Downlink OVSF codes are orthogonal until neighbour cells and multipath break them; uplink users are asynchronous and never orthogonal, so they are noise to each other: capacity is soft and the cell breathes. A chip lasts only ~0.26 us, right at urban delay spread, so orthogonality dies at the receiver. ★ CDMA could go fast; it could not go wide cheaply. HSPA+ reached ~42 Mbps around 2010 by bonding carriers — just bonding pipes.
  • ★★★★ The counter-intuitive part: LTE itself has no circuit domain. A bearer is IP end to end, the MSC has no counterpart, and voice demotes to one service over IP — that is VoLTE (IMS signalling, speech on a guaranteed QCI=1 bearer). Yet most LTE networks launched in 2010-2013 without VoLTE: calls borrowed the 2G/3G circuit domain via CSFB, and SRVCC handed a live call back on leaving LTE. That is why early phones dropped to 3G on a call. ★ Second trap: the uplink is SC-FDMA, not OFDMA, to spare the handset amplifier.
  • ★★★ Two takeaways. First, 150 Mbps is a cell-shared theoretical peak printed in a UE category (Rel-8, 20 MHz + 2x2), not your own throughput; carrier aggregation bonds pipes and improves no single pipe's SINR. Second, TD-LTE is not a separate Chinese standard: the TDD frame structure has been in 3GPP since Rel-8 (frozen 2008), sharing one physical layer with FDD, differing only in duplexing. China pushed the TDD branch and built the largest such network (late 2010s, base stations in the millions) — neither separate nor first.

(5) 5G - splitting the network

Real 0.1-1 Gbps · 2020s

This architecture: ★★★ 5G's real change is splitting: gNB into CU/DU plus AAU, the core into service-based NFs (SBA), control plane from user plane, the UPF placeable at the edge with MEC beside it. Speed comes from more Hz and more spatial streams.

What broke in the previous stage

  • ★★★ Set the premise first: the 4G page only claimed two things - about 150 Mbps peak, and one central P-GW anchor. 4G's control and user planes were never one box (MME and S-GW/P-GW are drawn apart); what jams is the P-GW itself: session control and policy enforcement (PCEF) fused with forwarding, an anchor you can neither split nor move, so all traffic detours back to it to exit (Rel-14 CUPS splits it, barely deployed). Speed hit bottom too: bits per Hz on one link at a given SNR is near Shannon, so the levers left are more Hz, more streams, denser cells.
  • ★★★★ The counterintuitive part: latency is not won on the radio. Fibre costs 0.5 ms per 100 km one way, so a detour to an anchor 500 km away eats ~5 ms round trip - no air-interface trick pays that back. 5G's answer is structural: control and user plane fully separated, the UPF an element you may place where you like, configured remotely by the SMF over N4 (PFCP), pushed down to a city or campus with MEC beside it and local traffic breaking out at N6. Slicing is the same story: isolation at the core and end to end, not a radio latency trick.
  • ★★★★ Three judgements you can use today. (1) Ask SA or NSA: in NSA (China uses Option 3x) 5G is only a secondary node under LTE, the core still EPC - throughput is real, but nothing on the 5GC side exists: no SBA, no end-to-end slicing, no VoNR. (2) No mmWave is still real 5G: FR1 and FR2 both count; mainland public networks had no commercial mmWave through end-2025, running on mid-band n41/n78, with sites in the millions. (3) 5G is not always greener: energy per bit fell, but a loaded 64T64R AAU draws kilowatts and sites are denser, so the power bill rises.

(6) 6G - Study

Target 50-200 Gbps · 2030s

This architecture: ★★ No 6G standard yet, only a framework: ITU-R IMT-2030 (M.2160, Nov 2023) gives scenarios and capability ranges it calls research targets; 3GPP has studied 6G in Rel-20 since 2025, first spec expected near 2029.

What broke in the previous stage

  • ★ First the premise: the 5G page has no satellite, AI or sensing box - not an omission; in 5G none of the three was part of the network structure. 5G's limit is not speed either (that page shows 0.1-1 Gbps measured): on its diagram mmWave is a separate small-cell layer, a hotspot only, while 3.5 GHz mid-band stays the coverage layer - and mid-band efficiency already nears the Shannon limit, so more MIMO layers or new coding add almost nothing. ★★ Capacity can only grow via new spectrum and spatial reuse. That is what forces a new generation.
  • ★★ The counter-intuitive part: 6G does not move higher. THz spectrum exists (WRC-19 identified ~137 GHz between 275 and 450 GHz), but 300 GHz costs ~39 dB more than 3.5 GHz at equal range, water-vapour absorption alone runs tens of dB per km in clear air and caps links under ~100 m; amplifiers give tens of mW, and ADC power roughly doubles per extra bit. THz fits short links of metres to tens of metres, plus sensing. So 6G steps down to upper mid-band FR3 (focus 7-15 GHz): existing sites, usable bandwidth.
  • ★★ Two things you can use now. One: satellites add coverage, not capacity - in direct-to-device form (operator low bands, tens of MHz) a beam spans hundreds of km sharing tens of Mbit/s, about one terrestrial cell, not comparable to a Ka-band dish. Two: peak is not experience - M.2160's 50-200 Gbit/s peak sits two orders above the 300-500 Mbit/s experienced rate. ★ And 6G not rebuilding the network means sites, bands and the 5GC architecture are reusable; the air interface is still new radio, not a software upgrade.

(7) Starlink - a different axis

Measured RTT 25-60 ms · 2020s

This architecture: ★★★ Not a 7th generation but a parallel axis to 4G/5G: the access point moves to 550 km, thousands of satellites plus laser crosslinks cover ocean and polar sites where no tower can stand. Most payloads are transponders; scheduling stays at the ground gateway, and capacity per km2 is 3-4 orders below terrestrial.

What broke in the previous stage

  • ★★★ Order first: this is not a 7th generation but an axis running parallel to 4G/5G - the NTN direct-to-phone satellite in the 6G stage is this axis being absorbed back into cellular, and only there does cellular write towerless cases into the standard. ★★ Every generation from 1G to 5G assumes one thing: a base station with site permit, backhaul and power inside the coverage area. Higher bands mean smaller cells and denser sites, so site dependence only grows. Where no tower can stand - ocean, desert, polar - walking from 1G to 5G still yields 0 bit/s.
  • ★★★ The usual mistake is what actually went up: for Ku broadband the satellite is basically a transponder, with modem and scheduler still at the ground gateway; only direct-to-cell truly flies an LTE eNB. ★★ Second trap is which latency you quote: GEO is ~477 ms of pure round-trip propagation at best, commonly 600 ms+ measured; 550 km leaves ~7 ms, with 25-60 ms measured end-to-end - a ~65x propagation gap but only 10-20x in experience. ★ Geometry's price: a satellite passes in minutes, and the terminal may be reassigned to another satellite or beam as often as every 15 s - handover reversed, the user still, the station flying.
  • ★★★ One rule to use right away: keep plan, ITU filing and in-orbit apart. SatNet's GW ~13,000 comes from a 2020 ITU filing, Qianfan's ~15,000 is mostly a company phased plan; as of Sep 2026 Qianfan is on the order of a hundred in orbit, GW a few dozen, Starlink seven-to-eight thousand. ★ The launch race comes from ITU rules: priority follows the filing date, no ownership is assigned, and milestones apply (first satellite in use within 7 years). ★★ But do not read it as replacing terrestrial: one beam reuses spectrum only every tens of km, so capacity per km2 is 3-4 orders lower.