Ferkans — Interactive Telecom Tutor

Making Cell-Free Real

The theoretical framework of §§1-4 produces dramatic performance numbers: 35% throughput gain, cm-level coverage uniformity, $10^{-13}$ BER at high mobility. But these assume: (i) sufficient fronthaul bandwidth, (ii) sub-microsecond AP synchronization, (iii) CPU compute matching the $L \times K$ scale. Each of these is a non-trivial engineering challenge at deployment scale. This section lays out the practical considerations that determine whether cell-free OTFS is actually deployable in a given context.

,

Definition:
Fronthaul Bandwidth Requirements

Cell-free OTFS fronthaul bandwidth per AP per frame:

Raw received signal (option 1, worst case): $N_a \cdot MN \cdot \text{bit width}$ . For $N_a = 4$ , $MN = 4096$ , 16-bit samples: 32 KB per frame.
Channel estimates only (option 2, standard): $P \cdot 4 \cdot \text{bit width}$ per UE in cluster. For $P = 8$ , $L_k = 6$ UEs: 768 B per frame per AP.
Precoded symbols (option 3, minimal): $MN \cdot N_a \cdot \text{bit width}$ after per-AP conjugate BF. Same size as raw, but content is data-bearing.

At 100 Hz frame rate:

Option 1: 3.2 MB/s per AP.
Option 2: 77 kB/s per AP.
Option 3: 3.2 MB/s per AP.

Aggregate across $L = 100$ APs:

Option 1: 320 MB/s. Supported by 10 GbE.
Option 2: 7.7 MB/s. Supported by 1 GbE.
Option 3: 320 MB/s. 10 GbE needed.

Practical choice: Option 2 (channel estimates) with optional Option 3 for high-precision scenarios.

,

Theorem: Fronthaul Bandwidth Scaling

The total fronthaul bandwidth for cell-free OTFS with $L$ APs, $K$ UEs, $N_a$ antennas per AP, $P$ paths per channel, $MN$ DD cells, and 100 Hz frame rate is $B_{\text{fronthaul}} \;=\; 100 \cdot L \cdot \min(L_k, K) \cdot P \cdot (4 \times 16 \text{ bits}).$ For urban-scale ( $L = 100$ , $L_k = 8$ , $K = 200$ , $P = 8$ ): $B \approx 5$ Gbps total fronthaul bandwidth. Distributed across $L = 100$ APs: 50 Mbps per AP average.

Optimization: user-centric clustering reduces the per-UE fronthaul bandwidth by factor $L_k/L$ . For $L_k = 8$ , $L = 100$ : $12\times$ reduction.

Fronthaul bandwidth scales as $L \cdot L_k \cdot P$ — linear in AP count, linear in cluster size, linear in path count. None of these scale with DD grid size ( $MN$ ) when we forward estimates only. This is the structural advantage of cell-free OTFS over cell-free OFDM: OFDM's per-subcarrier channel estimation requires $MN \cdot L_k$ coefficients per UE, $MN/P \sim 100\times$ more fronthaul.

Proof

Per-UE estimate size

$P$ paths, each with $(\tau, \nu, \theta, |a|)$ = 4 reals: $4P$ per UE per AP.

Per-AP aggregation

UEs in cluster: $L_k$ . Bits per estimate: $4 \times 16 = 64$ . Per AP per frame: $L_k \cdot 4P \cdot 64$ bits.

Total

$\times L$ APs, $\times 100$ Hz frame rate: $100 \cdot L \cdot L_k \cdot 256 P$ bits/s. $\blacksquare$

Definition:
Synchronization Requirements

Cell-free OTFS requires timing and phase synchronization across APs:

Timing sync (10 ns class): all APs must sample the OTFS frame at the same instants. Error $\epsilon_t$ translates to an effective inter-AP delay of $\epsilon_t$ — causes ISI if exceeding $\Delta\tau = 1/W$ .

Phase sync (1° class): conjugate BF requires coherent combining at the UE. Phase error $\delta\phi$ across APs causes $\cos(\delta\phi)$ signal loss. For $< 0.5$ dB loss: $\delta\phi < 15°$ .

Frequency sync: carrier frequencies across APs must match to $\sim 100$ Hz (for $\delta\nu < 1/T$ ). Usually achieved via shared frequency reference (GNSS-disciplined oscillator).

Synchronization methods:

GNSS-PPS + GNSS-disciplined oscillator: 50 ns timing, 1° phase, 100 Hz frequency. Standard for outdoor deployments.
PTP-1588v2 over fiber: 10 ns timing, 0.5° phase. For indoor cell-free or high-density deployments.
Hybrid GNSS + PTP: outdoor APs use GNSS; indoor/shielded use PTP fed from GNSS anchor.

,

Example: Sync Budget for Urban Cell-Free OTFS

Urban cell-free OTFS at 28 GHz, $W = 100$ MHz, $T = 16$ ms. Derive sync budget requirements.

Solution

Timing

$\Delta\tau = 1/W = 10$ ns. Timing error $\leq \Delta\tau/10 = 1$ ns for negligible ISI. Tight — GNSS-PPS at $\sim 50$ ns is insufficient. Need PTP or two-way time transfer.

Phase

For $< 0.5$ dB loss: $\delta\phi < 15°$ . At 28 GHz with $T = 16$ ms: relative oscillator drift over frame $< 15°/360° \cdot (1/T) = 15/(360 \cdot 0.016) = 2.6$ Hz. Requires GNSS-disciplined oscillator.

Frequency

$\delta\nu < 1/T = 62.5$ Hz. Much looser than phase. GNSS disciplining handles it.

Budget

Achievable with commercial GNSS + PTP hybrid hardware. Deployed in industrial cell-free installations. Cost: $\sim \$500$ per AP for sync hardware.

Definition:
Central Processing Unit (CPU) Architecture

The cell-free CPU architecture:

Input layer: receives per-AP channel estimates over fronthaul. Aggregates into $L \times K$ channel database.
Processing layer: computes precoders, performs resource allocation, monitors link quality.
- Per-UE precoder: $\mathcal{O}(P + N_a \cdot L_k)$ ops.
- Resource allocation: solves multi-user proportional-fairness or MMSE problem. $\mathcal{O}(L K)$ per iteration.
Output layer: forwards precoder vectors to APs (option 3 fronthaul), or forwards data symbols (option 1).

Deployment form factors:

Edge server (local, per-site): ~2U server, $\sim 10$ GFLOPS/core, 10-20 cores. Suitable for $L \leq 50$ .
Cloud CPU (remote, multi-site): GPU cluster. $\sim 100$ GFLOPS. $L \leq 500$ .
Hierarchical: local edge CPUs + regional cloud coordination. Standard for 6G O-RAN architectures.

Fronthaul Bandwidth vs Deployment Scale

Plot total fronthaul bandwidth as a function of $L$ for different options (raw, channel estimates, precoded symbols). Overlay capabilities of commercial fronthaul hardware.

Parameters

Max

L

200

K

200

N_a

4

Theorem: Cell-Free OTFS Scaling Limits

Cell-free OTFS is deployable up to scales limited by:

Fronthaul: total bandwidth $\leq 400$ Gbps (10 GbE × $L$ links). Supports $L \leq 1000$ APs with channel-estimate fronthaul.
CPU compute: per-frame ops $\leq 10^{10}$ . Supports $L K \leq 10^6$ user-AP pairs.
Synchronization: maintains quality for $L \leq 500$ APs with PTP; $L \leq 5000$ with GNSS + periodic calibration.
Pilot contamination: for 35% gain, requires $L/K \geq 5$ . Above $K = 1000$ : needs $L = 5000+$ APs.

Practical deployment scales:

Urban hot spot: $L = 50$ - $100$ , $K = 100$ - $200$ .
Dense urban: $L = 500$ - $1000$ , $K = 1000$ - $2000$ .
National-scale (6G): $L = 10^6$ APs. Possible, but requires hierarchical CPU + AI-based resource allocation.

The binding constraint depends on scale. For modest deployments ( $L < 100$ ), single edge CPU handles it with standard eCPRI fronthaul. For large deployments, hierarchical architecture distributes load. National-scale deployments need AI-based coordination — beyond current research but within 6G roadmap.

Proof

Fronthaul budget

10 GbE per AP × $L$ = $10 L$ Gbps. Useable fraction: 50%. Channel-est fronthaul: $\sim 50$ Mbps per AP. Maximum $L \sim 1000$ with 10 GbE per link.

CPU bandwidth

$L \times K$ channel matrix, plus precoder computation. Per-frame ops: $L K N_a + L K P$ . At 100 Hz: $10^8 L K$ ops/s. $10^{10}$ ops/s CPU: $L K \leq 10^2$ per-second average, or $L K \leq 10^6$ per-frame.

Sync

PTP cascade depth limits sync. Each hop adds 10 ns jitter. $L \leq 500$ : 2-hop PTP works. Beyond: GNSS needed.

Pilot contamination

For 35% gain: $L/K \geq 5$ . Scale to $K = 1000$ : need $L = 5000+$ . $\blacksquare$

🔧Engineering Note

O-RAN and eCPRI Integration

Cell-free OTFS integrates with the O-RAN (Open RAN) architecture:

AP ↔ CPU fronthaul: eCPRI 7.2 split. AP handles RF + sampling; CPU handles everything else including channel estimation.
AP ↔ CPU bandwidth: 10-25 GbE per AP. Standard for 5G deployments.
Synchronization: PTP-1588v2 over Ethernet. GNSS backup.
AI/ML integration: CPU can run AI models for resource allocation, user clustering, prediction. Part of 6G O-RAN RIC.

Cost: $\sim \$10$ k- $50$ k per AP fully provisioned (antenna, RF, fronthaul gear). At $L = 100$ : $1$ - $5$ M per site — comparable to cellular base station.

Deployment timeline: O-RAN cell-free prototypes in labs 2024- 2026. Commercial trials 2026-2028. Mass deployment 2028+ with 6G standardization.

Practical Constraints

•
eCPRI 7.2: standard cell-free fronthaul split
•
PTP-1588v2 for sub-microsecond sync
•
$10-50$ k per AP fully deployed
•
Commercial: 2028+ with 6G

Cell-Free OTFS Architecture and Data Flow

Animation of the cell-free OTFS architecture: distributed APs, UE, CPU, fronthaul. Data flow: pilots from UE → per-AP DD channel estimation → forwarded to CPU → aggregated → conjugate beamforming at APs → coherent signal at UE. The DD-domain processing happens at each AP locally, coordinated only through channel estimates at the CPU.

Why This Matters: Chapter 18: The Ultimate Cell-Free — LEO Constellation

Cell-free OTFS on the ground scales from 50 to 1000 APs over 1-100 km² areas. The next step — a LEO satellite constellation — is a cell-free network in the sky, with hundreds to thousands of satellites coordinating over the entire Earth. Chapter 18 (Buzzi- Caire-Colavolpe CommIT contribution) extends the cell-free framework to the orbital scale, where Doppler reaches ± 50 kHz and OFDM cannot compete.

Deployment: Fronthaul, Synchronization, Scaling