The 5G NR stack splits into three layers, and the engineering interest is almost entirely in where each function physically lands. O-RAN's functional splits exist because the answer changes with fronthaul bandwidth, latency budget and how much compute you are willing to put at the top of a mast.
What each layer is responsible for, and where it is typically implemented.
| Layer | Responsibility | Typical implementation |
|---|---|---|
| PHY (Layer 1) | Modulation, coding, rate matching, MIMO precoding, reference signals, FFT and beamforming weights | Hardware — FPGA or ASIC accelerator, with the digital front end alongside |
| MAC (Layer 2) | Scheduling, HARQ, multiplexing, random access, beam management decisions | Software on a real-time core, with hardware assist for HARQ |
| RLC (Layer 2) | Segmentation and reassembly, ARQ retransmission, duplicate detection | Software |
| PDCP (Layer 2) | Header compression, ciphering and integrity protection, reordering, duplication for reliability | Software, with cryptographic offload to hardware |
| SDAP (Layer 2) | Quality-of-service flow to data radio bearer mapping | Software |
| RRC (Layer 3) | Connection establishment, mobility, measurement configuration, bearer management | Software, control plane only |
| NAS (above Layer 3) | Registration, session management, authentication with the core | Software, terminates in the core network |
The layer boundaries are specified; the hardware and software split is a design decision, and it is the decision that determines cost, power and how quickly a product can follow a standards revision.
The split point trades fronthaul bandwidth against how much processing must sit at the radio site.
| Split | Cut point | Fronthaul demand | Consequence |
|---|---|---|---|
| Option 1 | Between RRC and PDCP | Low | Almost everything at the radio site; little centralisation benefit |
| Option 2 | Between PDCP and RLC | Low | The CU/DU split in practice. Latency-tolerant, widely deployed |
| Option 6 | Between MAC and PHY | Moderate | Scheduling centralised, PHY distributed. Tight latency requirement |
| Option 7.2x | Within PHY, after precoding | High | The O-RAN choice. Radio unit stays comparatively simple; fronthaul carries frequency-domain samples |
| Option 8 | Between PHY and RF | Very high | Raw time-domain samples. Simplest radio unit, fronthaul cost usually prohibitive at scale |
Option 7.2x is where O-RAN settled because it keeps the radio unit tractable and the fronthaul affordable at massive MIMO antenna counts. Split choice is not a preference; it is arithmetic on antenna count, bandwidth and fibre availability.
Splitting the stack creates three network functions with different real-time obligations, different siting and different hardware. A programme that says “O-RAN” without saying which node is not yet a specification.
| Node | What it runs | Where it sits and why |
|---|---|---|
| Radio unit (RU) | Lower physical layer and the RF chain, including the digital front end | At the antenna. Hard real-time, thermally constrained, replicated across every site, so unit cost and amplifier efficiency dominate |
| Distributed unit (DU) | Upper physical layer, MAC and RLC — the scheduler lives here | Close to the radios, because the hybrid ARQ and scheduling deadlines are measured in hundreds of microseconds. Usually accelerated hardware alongside general-purpose compute |
| Central unit (CU) | PDCP, SDAP and RRC — connection, mobility and security | Pooled, often virtualised, serving many distributed units. Millisecond-scale rather than symbol-scale, so it tolerates being far away |
Fronthaul connects RU to DU and carries the tightest timing in the network. Midhaul connects DU to CU and is comparatively relaxed. Backhaul carries the CU to the core.
5G NR scales subcarrier spacing across 15, 30, 60 and 120 kHz, and slot duration scales inversely — 1 ms at 15 kHz down to 125 microseconds at 120 kHz. Higher numerology buys lower latency and better phase-noise tolerance at millimetre wave, and it costs proportionally tighter processing deadlines.
That deadline is the design constraint. At 120 kHz spacing the entire receive chain — FFT, channel estimation, equalisation, demodulation, decoding — has to complete inside a fraction of 125 microseconds, because HARQ feedback is due. Miss it and throughput collapses, not gracefully.
In time-division duplex the same clock governs the transmit-to-receive turnaround. Guard periods are set by propagation delay and by how quickly the front end can switch, and getting that wrong shows up as interference to a neighbour rather than as a fault on your own equipment.
PHY at Layer 1; MAC, RLC, PDCP and SDAP at Layer 2; RRC at Layer 3; and NAS above it terminating in the core network. PHY is typically hardware, Layer 2 and above typically software with cryptographic and HARQ offload.
A functional split inside the physical layer, after precoding, so the radio unit handles the lower PHY and RF while the distributed unit handles the upper PHY and above. It is O-RAN's choice because it keeps the radio unit tractable while keeping fronthaul bandwidth affordable at high antenna counts.
Because slot duration scales inversely with it. At 120 kHz spacing a slot is 125 microseconds, and the entire receive chain must complete within a fraction of that for HARQ feedback to be on time. The numerology sets the processing deadline.
PHY is hardware because of the deadline. MAC and above are software for flexibility, with hardware assist for HARQ and cryptography. Exactly where the boundary sits is a design decision that determines cost, power and how fast a product can track a standards revision.
Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.