Fine-Grained Co-OFDMA

• Fine-Grained Co-OFDMA is a method enabling multi-access-point coordination at the resource unit level within a single 20 MHz Wi-Fi channel, optimizing spectrum usage in dense networks.
• It leverages a wired fiber backhaul with White Rabbit for precise time synchronization and carrier-frequency offset compensation, ensuring Wi-Fi 6 compliance and reliable performance.
• Experimental results demonstrate reduced latency, with inter-AP CFO differences of 27.34 Hz and timing deviations within ±48 ns, improving short-packet transmission efficiency.

Fine-grained coordinated orthogonal frequency division multiple access (Co-OFDMA) denotes multi-access-point coordination at resource-unit granularity within a single $20\,\text{MHz}$ Wi-Fi channel, so that multiple APs transmit simultaneously on disjoint RUs under a common HE-MU frame structure rather than coordinating only at whole-channel granularity. In dense deployments, the motivation is to reduce the latency penalties of contention-based access, especially when multiple APs serve stations in overlapping coverage and may be hidden from each other. A practical realization using a wired fiber backhaul, openwifi, and White Rabbit demonstrates that such coordination can remain Wi-Fi 6 compliant while achieving carrier-frequency-offset pre-compensation and time synchronization that exceed the cited wireless-standard requirements; the reported residual inter-AP CFO difference is $27.34\,\text{Hz}$, timing deviation is within $\pm 48\,\text{ns}$, and the received constellation quality of the joint Co-OFDMA frame is better than that of individual AP transmissions (Havinga et al., 14 Jul 2025).

1. Conceptual basis and problem setting

Coordinated OFDMA is proposed for Wi-Fi 8 as a remedy for the latency degradation that contention-based systems experience when multiple devices compete for the same resources. The fine-grained variant addresses coordination not merely across separate channels, but within a single $20\,\text{MHz}$ channel at RU level. In the demonstrated setting, multiple APs share the spectrum more efficiently by agreeing which RUs each AP occupies, allowing simultaneous transmission without mutual interference. The paper argues that fine-grained allocation within $20\,\text{MHz}$ is often regarded as impractical because over-the-air scheduling incurs substantial overhead and because physical-layer signaling becomes more complex; the central claim is that a wired backhaul changes that feasibility assessment (Havinga et al., 14 Jul 2025).

The operational scenario is a dense deployment in which APs may not be within each other’s range yet still serve stations in overlapping coverage. Without coordination, simultaneous transmissions collide. A wired backhaul eliminates the need for AP-to-AP over-the-air discovery, buffer-state exchange, timing negotiation, and synchronization, and therefore permits coordination even when APs are hidden from one another. The same infrastructure is presented as a basis for broader multi-AP coordination, including coordinated beamforming and joint transmission.

The “fine-grained” qualifier is not incidental. Restricting coordination to whole $20\,\text{MHz}$ channels is described as too coarse for many low-rate, low-latency flows, because small payloads would occupy wider channels than necessary, increasing preamble overhead, noise exposure, and power dilution. RU-level coordination instead supports low-rate traffic more efficiently and can mitigate narrowband cross-technology interference.

2. Latency rationale and efficiency of sub-$20\,\text{MHz}$ coordination

The paper’s latency argument is illustrated with a concrete two-AP downlink example. For a $500$-byte packet per AP at MCS 7 over $20\,\text{MHz}$, a single-user transmission requires $104\,\mu\text{s}$ airtime. If hidden-node protection is implemented through RTS/CTS, the schedule accumulates a DIFS of $27.34\,\text{Hz}$0, an RTS of $27.34\,\text{Hz}$1, a CTS of $27.34\,\text{Hz}$2, SIFS gaps totaling $27.34\,\text{Hz}$3, and contention/backoff overhead, yielding a best-case total of $27.34\,\text{Hz}$4 for the two packets. Under Co-OFDMA, one AP contends, sends a trigger over the fiber link with a modeled $27.34\,\text{Hz}$5 latency, and both APs transmit a two-user MU frame simultaneously; using two $27.34\,\text{Hz}$6-tone RUs, the concurrent transmission completes in $27.34\,\text{Hz}$7 (Havinga et al., 14 Jul 2025).

This comparison isolates the principal systems implication: the backhaul shifts coordination from airtime-consuming wireless control exchange to deterministic wired signaling. The result is not only reduced scheduling overhead, but also a coordination mechanism that remains available when APs cannot decode one another’s frames. The fine-grained formulation is therefore tied directly to short-packet efficiency, hidden-node mitigation, and reuse of a single $27.34\,\text{Hz}$8 channel by multiple APs.

A common misunderstanding is that Co-OFDMA mainly benefits large aggregated flows. The reported example instead emphasizes short downlink packets, where conventional protection overhead is proportionally large. In that regime, RU-level sharing rather than channel-level partitioning is the salient gain.

3. Fiber-backhaul architecture: openwifi, White Rabbit, and the control path

The implementation combines openwifi and White Rabbit. openwifi provides the open-source Wi-Fi transceiver and MAC stack on SDR hardware, with the low PHY and low-MAC in FPGA programmable logic and the upper MAC on the ARM processing system. The prototype builds on earlier work that extended openwifi with IEEE 802.11ax OFDMA support. White Rabbit, originally from CERN, supplies synchronization over fiber through three components: PTP for time synchronization, SyncE for frequency syntonization, and DDMTD for phase detection. The prototype uses a “Light Rabbit” approach that runs White Rabbit on an AMD ZCU102 SoC without extra dedicated hardware (Havinga et al., 14 Jul 2025).

The physical setup uses two ZCU102-based APs directly linked by fiber through the onboard SFP transceiver. The paper notes that a larger White Rabbit network could also be built through a WR switch. White Rabbit provides both the syntonized reference clock and the deterministic low-latency transport needed for timing the trigger and estimating carrier offset.

The control path is intentionally minimal. Each AP continues to run its own CSMA/CA. When one AP wins channel access, it sends a trigger over the fiber link using the WR Streamers module, which provides low-latency Ethernet-based transport. The second AP receives the trigger and begins baseband processing after the calibrated transmission delay. There is also a safeguard: if the second AP senses that the medium is busy, it may ignore the trigger and withdraw, analogous to the “CS required” field in Wi-Fi 6 uplink OFDMA. The architecture therefore preserves local contention behavior while relocating only the critical coordination step to the backhaul.

4. Frequency alignment and time synchronization

Carrier-frequency-offset handling is a central technical issue because the AD9361 RF front end uses a $27.34\,\text{Hz}$9 crystal with $\pm 48\,\text{ns}$0 stability. At the carrier frequency of Wi-Fi channel 1, $\pm 48\,\text{ns}$1, that stability can induce up to $\pm 48\,\text{ns}$2 CFO. Compensation is performed by tuning the AD9361 DCXO, whose worst-case tuning resolution is $\pm 48\,\text{ns}$3, corresponding to $\pm 48\,\text{ns}$4 at $\pm 48\,\text{ns}$5. The design uses the White Rabbit-syntonized $\pm 48\,\text{ns}$6 reference clock inside the FPGA to estimate the local-oscillator frequency relative to the White Rabbit reference. The CFO Estimation module counts the number of LO clock cycles over a known number of $\pm 48\,\text{ns}$7 reference cycles and compares the observed count with the expected count. To keep the measurement error below the DCXO tuning step, the measurement window is set to $\pm 48\,\text{ns}$8, or $\pm 48\,\text{ns}$9 million $20\,\text{MHz}$0 cycles, corresponding to $20\,\text{MHz}$1 million expected $20\,\text{MHz}$2 rising edges and a $20\,\text{MHz}$3 measurement resolution; at $20\,\text{MHz}$4, the maximum measurement error is $20\,\text{MHz}$5. The measured count difference is written to a register shared with the ARM processor, and a driver task periodically sends SPI commands to adjust the AD9361 DCXO, thereby avoiding PLL relock during transmission (Havinga et al., 14 Jul 2025).

For time synchronization, the adopted constraint matches IEEE 802.11ax uplink OFDMA timing: transmissions from different APs should begin within the $20\,\text{MHz}$6 guard interval, or within $20\,\text{MHz}$7 relative to one another. The reason given is that, as long as the relative delay remains within the guard interval, cyclic-prefix insertion prevents inter-symbol interference and preserves subcarrier orthogonality. Round-trip trigger latency is measured by having one AP send a trigger and the other immediately respond, allowing inference of one-way timing behavior. Across $20\,\text{MHz}$8 triggers from AP1 and $20\,\text{MHz}$9 from AP2, the measured round-trip time averages $20\,\text{MHz}$0 with a standard deviation of $20\,\text{MHz}$1. With latency quantized to the White Rabbit system clock at $20\,\text{MHz}$2, the maximum deviation is three clock cycles, or $20\,\text{MHz}$3, which lies well within the $20\,\text{MHz}$4 requirement. The inferred one-way latency is about $20\,\text{MHz}$5, which is below the IEEE 802.11 receive-to-transmit turnaround requirement of $20\,\text{MHz}$6. The paper identifies WR Streamers and a native fiber link as possible routes to further latency reduction.

These measurements distinguish between frequency alignment, syntonization, and the stronger condition of phase coherency. The demonstrated system achieves the first two to the degree needed for Co-OFDMA; the authors explicitly note that joint transmission would require phase coherency.

5. RU-level packet construction and standards compatibility

Fine-grained Co-OFDMA is realized through the $20\,\text{MHz}$7 HE-MU packet structure with RU-level separation. The implementation retains Wi-Fi 6 compliance so that the resulting frame remains standard-compliant and decodable by stations. Both APs construct the same full-bandwidth preamble through HE-SIG-B, which implies that all fields affecting decoding must be agreed in advance. The paper specifically identifies the L-Length field in L-SIG, the BSS color in HE-SIG-A, the STA-ID assignments, and the RU allocation map in HE-SIG-B as shared state that must be consistent across APs (Havinga et al., 14 Jul 2025).

This requirement has direct protocol consequences. Because stations normally use BSS color to decide whether to decode a frame, APs participating in Co-OFDMA must share the same BSS color. Likewise, STA-ID collisions across APs must be avoided by assigning different STA-IDs in their respective BSSs. After the shared preamble, each AP transmits only on its assigned RU and zeros out the subcarriers belonging to the other AP. The paper’s exemplar within one $20\,\text{MHz}$8 channel uses two $20\,\text{MHz}$9-tone RUs so that two APs can share the channel simultaneously. It also notes that standard Wi-Fi 6 OFDMA allows up to nine users per $20\,\text{MHz}$0 using RU sizes of $20\,\text{MHz}$1, $20\,\text{MHz}$2, or $20\,\text{MHz}$3 subcarriers. Power in each occupied RU is scaled to account for the smaller number of active subcarriers, increasing power spectral density similarly to uplink OFDMA.

The reported spectrum observations reinforce the packet-format argument. When AP1 is set to channel 1 and AP2 to channel 6, both APs transmit an identical full-bandwidth preamble simultaneously, after which each occupies only its own $20\,\text{MHz}$4-tone RU. When both APs are set to channel 1, the full-bandwidth preambles overlap and the APs occupy distinct portions of the same $20\,\text{MHz}$5 channel. A transient artifact appears at the start of each OFDM symbol because cyclic-prefix insertion introduces a time-domain discontinuity.

6. Experimental characterization

The experimental platform uses two AMD ZCU102 boards interconnected by roughly $20\,\text{MHz}$6 of fiber. Their RF outputs are combined through a power combiner and fed into an R&S CMW270 wireless connectivity tester; for spectral visualization, the combined output is also routed to an Anritsu MS2690A spectrum analyzer. The measurements cover CFO, timing, spectrum alignment, and EVM (Havinga et al., 14 Jul 2025).

The aggregate results are summarized below.

Quantity	Context	Reported value
Initial CFO	Before tuning	AP1 about $20\,\text{MHz}$7, AP2 about $20\,\text{MHz}$8
Residual inter-AP CFO difference	Wireless tester, after tuning	$20\,\text{MHz}$9
Residual CFO limit	Referenced from IEEE 802.11ax uplink OFDMA requirements	$20\,\text{MHz}$0
Timing deviation	Quantized WR-latency bound	$20\,\text{MHz}$1
Timing requirement	IEEE 802.11ax uplink OFDMA timing target	$20\,\text{MHz}$2
Average EVM	Individual AP transmissions	$20\,\text{MHz}$3 for AP1, $20\,\text{MHz}$4 for AP2
Average EVM	Joint Co-OFDMA transmission	$20\,\text{MHz}$5
Burst power	Individual AP transmissions	about $20\,\text{MHz}$6 and $20\,\text{MHz}$7
Burst power	Joint Co-OFDMA transmission	$20\,\text{MHz}$8

The temporal behavior of CFO tuning is also reported. With tuning enabled, both APs are tuned to the White Rabbit reference after about $20\,\text{MHz}$9, and the CFO difference remains within roughly $500$0, with a few tuning events per minute. When tuning is disabled, oscillator drift causes the difference to exceed the $500$1 residual CFO limit cited from IEEE 802.11ax uplink OFDMA requirements. As measured by the wireless tester, the average CFO is $500$2 for AP1 and $500$3 for AP2, leaving the $500$4 residual difference.

For signal quality, the paper reports EVM and burst power over $500$5 individual MU OFDMA packets per AP and $500$6 joint Co-OFDMA packets. The joint transmission improves average EVM by about $500$7 relative to the individual transmissions, while burst power increases to a value close to the expected $500$8 gain from combining two APs’ power over disjoint subcarriers. The paper attributes the smaller-than-$500$9 EVM gain to transmitter noise becoming dominant at higher signal levels.

7. Scope, misconceptions, and relation to adjacent “fine-grained” OFDMA literature

Several distinctions are necessary for precise use of the term. First, fine-grained Co-OFDMA is not identical to whole-channel coordination. Its defining granularity is RU-level sharing within one $20\,\text{MHz}$0 channel. Second, it is not synonymous with over-the-air multi-AP scheduling: the demonstrated system specifically uses fiber backhaul to remove the need for AP-to-AP over-the-air scheduling and synchronization. Third, it is not yet joint transmission. The paper presents Co-OFDMA as a stepping stone toward coordinated beamforming and eventually joint transmission, but explicitly states that joint transmission would require phase coherency rather than only the reported time and frequency alignment (Havinga et al., 14 Jul 2025).

The phrase “fine-grained” also appears in adjacent OFDMA and IFDMA work with different meanings. In "Flexible Subcarrier Allocation for Interleaved Frequency Division Multiple Access" (Shao et al., 2020), the issue is single-system subcarrier-allocation flexibility: bit-reversal mapping converts evenly spaced IFDMA allocation into contiguous-bin filling, and multi-stream IFDMA removes the divisor constraint so that any positive integer $20\,\text{MHz}$1 can be supported. That work shows that, in the synchronous scenario, IFDMA can achieve the same level of flexibility and granularity as OFDMA in subcarrier allocation, and in the asynchronous scenario the blocking-probability gap to OFDMA is only $20\,\text{MHz}$2 at a moderate offered load. In "OFDMA-F$20\,\text{MHz}$3L: Federated Learning With Flexible Aggregation Over an OFDMA Air Interface" (Hu et al., 2023), “fine-grained” instead refers to per-round client, subchannel, and modulation selection under a delay budget, with an almost surely optimal “winner-takes-all” assignment. Taken together, these usages indicate that “fine-grained” in OFDMA-adjacent research is a granularity descriptor rather than a single standardized mechanism.

A plausible implication is that fine-grained Co-OFDMA should be understood as one layer in a broader hierarchy of resource-allocation granularity. In the Wi-Fi multi-AP setting, the relevant granularity is inter-AP RU partitioning inside one channel; in IFDMA, it is subcarrier and stream composition; in OFDMA-enabled federated learning, it is per-round subchannel and modulation selection. The significance of the Wi-Fi realization lies in showing that backhaul-enabled coordination can make the multi-AP RU-level case practical while remaining standards-compatible and compatible with future extensions toward more advanced coordinated transmission modes.