5G NR PHY Architecture for RAN Architects — The 1-Page Mental Model
RAN architects don't need to know the bit-level mechanics of LDPC. They need to know where layer-1 boundaries fall, what crosses each interface, and what the trade-offs are. Here's the 1-page mental model.
The Three Boundary Decisions
Designing a 5G RAN means picking three boundaries:
- CU vs DU — F1 interface. PDCP/RRC vs RLC/MAC/PHY split.
- DU vs RU — Open Fronthaul. Where in PHY do we cut?
- Centralized vs Distributed PHY — How much PHY runs in software vs dedicated hardware?
F1 — The CU/DU Boundary
F1 carries PDCP PDUs from CU to DU. Centralized CU (1 per region, 100s of DUs) enables:
- Centralized RRC, simplified mobility (CU coordinates handovers between DUs)
- Pooled CU compute — CU farms running Linux on COTS
- Standardized F1 interface (TS 38.470 series) — multi-vendor possible
Latency budget: F1 < 5 ms typical. Within geographic region, this is solvable with backhaul fiber.
Open Fronthaul — The DU/RU Boundary
O-RAN 7-2x split places the boundary inside the OFDM transmit chain. Frequency-domain precoded I/Q samples cross the fronthaul (much lower bandwidth than time-domain raw I/Q like CPRI 7-1).
Latency budget: < 100 μs one-way fronthaul typical. Some operators target < 50 μs for tight HARQ. Fiber distance ≤ 20 km is feasible.
Interface: eCPRI over Ethernet (typically 25G or 100G), with Time-Sensitive Networking (TSN) for sync. SyncE + PTP combined.
What Lives in DU vs RU
DU (Distributed Unit): RLC, MAC, scheduler, high PHY: LDPC encoding, scrambling, modulation, layer mapping, precoding. This is where the heavy compute is.
RU (Radio Unit): Low PHY: resource mapping, IFFT (or FFT for UL), CP insertion, RF front-end, antenna array. RU ships frequency-domain I/Q upstream (UL) or downstream (DL).
For massive MIMO: RU performs analog beam-steering (via phase-shifter network on the array); DU computes digital precoding matrix W.
Software vs Hardware PHY
Modern DUs run PHY in a mix:
- FPGA / accelerator card: LDPC decode, FFT/IFFT, channel estimation. Latency-critical, hardware-accelerated.
- x86 CPU + DPDK: Scheduler, MAC, RLC, parts of high PHY. Software-flexible.
- GPU (some vendors): CSI computation, AI/ML inference (beam prediction, etc.).
This mix is what enables vRAN — virtualized RAN where DU runs as software-defined instances on COTS servers + accelerator cards.
The RIC — RAN Intelligent Controller
O-RAN adds management plane: SMO + RIC.
- Near-RT RIC: 10ms-1s loop. Hosts xApps for: dynamic load balancing, traffic steering, beam steering optimization, anomaly detection.
- Non-RT RIC: > 1s loop. Hosts rApps for: long-term policy decisions, AI/ML model training, configuration management.
For RAN architects, this introduces a new layer of design: which optimizations are RIC-driven vs gNB-internal? Standard answer: RIC for policies that span multiple gNBs; gNB internal for slot-by-slot decisions.
The Architect's Trade-Off Table
| Choice | Centralized | Distributed |
|---|---|---|
| CU | Cheaper (pooled), simpler mobility | Lower latency to DU |
| DU | Pooled compute, lower TCO | Smaller fronthaul fiber span |
| PHY | vRAN agile, software-defined | Hardware-locked, max performance |
What This Means in Practice
Most 2026 deployments: distributed RU in cell tower, centralized DU farm in metro PoP (~5-20 km fiber to RU), centralized CU farm in regional DC (~50-100 km from DU). RIC near-RT runs at the CU farm; rApps centralized at SMO.
The full 5G NR PHY course covers the layer-1 details that flow over each interface — what's in the F1 PDU, what's in the eCPRI U-plane, how scheduling decisions reach the RU. 99 lessons, $29 lifetime.
Master the full 5G NR Physical Layer
This article covers one topic. The full course is 99 lessons across 6 modules — every channel, every algorithm in TS 38.211–215 — with cinematic animations.
Enroll · $29 lifetime →