Access Networks — DSL, Cable, FTTH, Cellular
How end hosts get on the Internet — the last-mile technologies, their bandwidth shapes, and where they're going.
Summary#
The access network is the last hop between an end host and the public Internet — the wire (or radio link) from your laptop, phone, or office to your ISP’s first router. The Internet’s core is fast and uniform; the access network is where most of the variance lives. Latency, bandwidth, asymmetry, and reliability are mostly determined here.
Four families dominate the consumer and small-business last mile: DSL (twisted-pair copper, telco origin), cable (coaxial, cable-TV origin), FTTH / FTTx (fiber, the modern default for new deployments), and cellular (3G/4G/5G, the only one that moves with you). Each has a different physical medium, a different shared-vs-dedicated structure, and a different bandwidth shape — and which one feeds your house determines, more than anything else, your real Internet experience.
Why it matters#
Most user-perceived latency lives in the access network. A round-trip from a residential cable modem to the ISP edge can be 5-20 ms before any inter-AS hops. Cellular adds another 10-50 ms of radio scheduling on top. By contrast, intra-datacenter latency is sub-millisecond. When you’re designing for “feels instant” you are mostly fighting the last mile.
It also explains asymmetric bandwidth, which is the default everywhere except FTTH. A typical cable plan is ~500 Mbps down / 20-50 Mbps up. DSL was even more skewed. This is a hardware choice driven by the original use case (downloading content) and the modulation/duplexing scheme. It matters because uploads — video calls, backups, P2P, content creation — saturate first and silently degrade everything else sharing the link (head-of-line on the cable modem’s uplink queue).
Access-network choice also drives reliability. FTTH outages are rare but long. Cable is reliable but congested. Cellular is everywhere but throttled. Enterprise access (dedicated fiber, MPLS) is what you buy when “reliable” means SLA-backed.
How it works#
The four major access technologies, each terminating at a different demarcation point:
home / phone ISP edge +----------+ +----------+DSL ---- | DSL modem | ---- twisted pair ---| DSLAM |---> aggregation -> core +----------+ +----------+
+----------+ +----------+Cable ---- | DOCSIS | ---- coax --------- | CMTS |---> aggregation -> core | modem | (shared trunk) +----------+ +----------+
+----------+ +----------+FTTH ---- | ONT (opt) | ---- | OLT (PON) |---> aggregation -> core +----------+ +----------+ fiber to home (passive splitter)
+----------+4G/5G ---- | UE (phone)| ---- radio (RAN) --> eNodeB/gNodeB --> EPC/5GC --> core +----------+DSL (Digital Subscriber Line)#
Runs over the same twisted-pair copper that carries POTS (plain old telephone service), using frequencies above the voice band. ADSL was asymmetric and capped around ~24 Mbps down / 3 Mbps up. VDSL2 with vectoring pushes ~100 Mbps. Distance-limited — performance falls sharply as the copper run from the DSLAM to the home exceeds 1-2 km. Mostly being decommissioned in favour of fiber.
Cable (DOCSIS)#
Cable plant was originally one-way broadcast TV over coax. DOCSIS added a return path. The plant is shared: many homes hang off one coax trunk that terminates at a Cable Modem Termination System (CMTS). Bandwidth is divided among everyone on the segment — peak-hour slowdown is the canonical complaint. DOCSIS 3.1 supports multi-gigabit downstream but uplinks remain modest (~50 Mbps typical, 1 Gbps on DOCSIS 4.0).
FTTH (Fiber to the Home)#
Fiber from the ISP all the way to the customer premises. Two flavours dominate:
- PON (Passive Optical Network) — one fiber from the OLT (in the central office) splits via a passive optical splitter to 32 or 64 customers. Each home has an ONT. Shared upstream timeslot, dedicated downstream wavelength. GPON gives 2.5 Gbps shared; XGS-PON 10 Gbps symmetric.
- Active Ethernet / point-to-point — dedicated fiber per customer. Premium pricing, deterministic bandwidth.
FTTH is the only family with truly symmetric multi-gigabit options at consumer prices. It’s what new builds default to.
Cellular (3G / 4G / 5G)#
Radio link from the user equipment (UE) to a base station (eNodeB on 4G, gNodeB on 5G), backhauled over fiber to the operator’s core network. Two distinguishing properties:
- Mobility — the radio resource control protocol handles handover between base stations, IP address survives via mobility anchors. None of the other access types support moving.
- Scheduled access — unlike Wi-Fi’s CSMA/CA, the base station scheduler assigns time and frequency resources to each UE every millisecond. Latency floor is the scheduling slot (
~1 mson 5G NR, more on 4G).
5G adds three big things: higher spectral efficiency (massive MIMO), millimetre-wave bands with multi-gigabit but short reach, and network slicing (logically separate networks on one RAN for different use cases — eMBB for phones, URLLC for low-latency industrial, mMTC for IoT).
Enterprise access#
Dedicated fiber (MetroE, dark fiber, EPL), MPLS VPNs, SD-WAN over multiple commodity links. Buys SLA, symmetric bandwidth, and burstable capacity. Costs orders of magnitude more per megabit but the consistency matters.
Variants and trade-offs#
Other axes:
- Symmetric vs asymmetric bandwidth — DSL, cable, and PON are asymmetric. Active Ethernet, XGS-PON, and 5G mid-band lean symmetric. Symmetry matters for upload-heavy workloads (cloud backup, P2P, video calls).
- Shared vs dedicated — cable and PON share the medium among neighbours; cellular shares the radio resource among everyone in the cell. Active Ethernet is dedicated. Saturation behaviour differs.
- Wired vs wireless — wireless is convenient and mobile but has higher and more variable latency, lower spectral efficiency under congestion, and is vulnerable to interference.
Why is upload bandwidth so much smaller?
Three reasons. Historical: DSL and cable plants were laid for downstream-heavy use cases (web, video, downloads). Physical: in cable (DOCSIS) the upstream uses lower frequencies where noise is worse and channels are narrower; the spectrum allocation is hard to undo without re-plant. Modulation: higher-order QAM works better when there’s a stable signal (downstream from one transmitter to many), worse when many transmitters share the uplink (asymmetric noise budgets). FTTH escapes all three — fiber doesn’t share spectrum with anything, and PON’s downstream and upstream wavelengths can both be widened.
When this is asked in interviews#
Asked mostly as context for performance and architecture questions, less as a memorisation test. “Why does my video call freeze every evening?” Answer: shared access medium (cable, cellular) plus everyone home from work; bufferbloat at the CMTS; uplink saturation from someone else’s backup.
Follow-ups to expect:
- “Why is upload so much slower than download?” — DOCSIS spectrum allocation, modulation differences, plant history.
- “What’s the difference between FTTH and FTTC?” — FTTH brings fiber all the way; FTTC (fiber to the cabinet) runs fiber to a street cabinet and copper for the last few hundred metres (essentially VDSL).
- “What does 5G actually give you?” — higher spectral efficiency, sub-6 GHz speed bump, mmWave for capacity hot spots, low scheduling latency, network slicing.
- “How does the carrier give my phone an IP address?” — when the UE attaches, the PDN gateway in the EPC/5GC assigns an IP (usually IPv6 with a per-UE prefix; IPv4 may be NATted). Mobility anchors keep the IP stable across cell handover.
Most relevant in roles touching infrastructure capacity planning, mobile backend, video/voice products, and CDN placement.
Related concepts#