Wi-Fi 7 (802.11be): MLO, 320 MHz, and the Latency Push

Overview

802.11be — certified by the Wi-Fi Alliance as Wi-Fi 7 — is the “Extremely High Throughput” amendment that turns the 6 GHz era into a multi-radio, multi-link discipline: channels double to 320 MHz, constellation density rises to 4096-QAM, and Multi-Link Operation (MLO) lets one client and one AP aggregate two or more physical radios into a single logical connection with per-packet link selection. The headline PHY rate is 46.1 Gbps, but the interview-worthy content is the arithmetic that shrinks it to real 5 Gbps throughput, the link types (STR vs. non-STR) that decide whether MLO actually aggregates, and the scheduling features (punctured preamble, R-TWT) that trade peak rate for bounded latency in AR/VR-class traffic. The generation-overview table lives on the Wi-Fi page; this page goes deep on the 802.11be mechanics themselves.

From 11ax to 11be: What Actually Changed

Wi-Fi 6 (802.11ax) built the efficiency toolkit — OFDMA, uplink scheduling via trigger frames, 1024-QAM, TWT power saving — and Wi-Fi 6E opened the 6 GHz band for it. Wi-Fi 7 keeps all of that and changes the physics and the topology:

Two changes are structural rather than incremental. Multiple RUs per user means one OFDMA transmission can give a single client, say, a 106+26 tone pair instead of forcing a choice between one big RU and nothing. And the 512-entry compressed block ack bitmap matters more than it looks: wide channels and MLO raise the natural aggregation window, and a 256-bit bitmap became the accounting bottleneck at 320 MHz with reordering across links.

Spectrum: 2.4/5/6 GHz and 320 MHz Channels

The 320 MHz channel exists because of 6 GHz. The band adds roughly 1200 MHz of spectrum (59 new 20 MHz channels in the 6E allocation), and only there can three non-overlapping 320 MHz channels fit contiguously; the 5 GHz band tops out at 160 MHz, and 2.4 GHz at 40 MHz. That asymmetry drives real deployments: a 320 MHz-capable AP dedicates 6 GHz radio capacity to a handful of nearby, capable clients, and everyone else lives at 160 MHz or below. Because 6 GHz is shared with incumbents, its use is mediated — by Automated Frequency Coordination (AFC) for standard-power operation in the US, and by Low Power Indoor restrictions in regions (the EU released the lower 500 MHz subset) — which is why channel plans and regulatory domains are part of any serious Wi-Fi 7 design conversation.

The corollary is MLO’s second job. With only three clean 320 MHz channels, contention and incumbents fragment the band in practice; MLO plus puncturing lets an AP assemble usable capacity from imperfect spectrum instead of cliff-dropping from 320 to 160 MHz whenever one 20 MHz slice is occupied.

4096-QAM and Throughput Math

Deriving the 46.1 Gbps

The headline number is arithmetic, not marketing, and deriving it is the interview exercise. Start from 802.11ax: at 160 MHz, 1024-QAM, 5/6 coding, the highest data rate is 9.6 Gbps across 8 spatial streams — 1.2 Gbps per stream. Wi-Fi 7 doubles the channel to 320 MHz (factor 2), raises the constellation to 4096-QAM, i.e. 12 bits per symbol instead of 10 (factor 1.2), and doubles the stream count to 16:

\[ R_{\text{stream}} = 1.2 \text{ Gbps} \times 2 \times 1.2 = 2.88 \text{ Gbps} \qquad R_{\text{PHY}} = 2.88 \times 16 = 46.1 \text{ Gbps} \]

A realistic flagship AP radio is 4x4 per band, so its 320 MHz ceiling is \( 2.88 \times 4 \approx 11.5 \) Gbps of PHY — and a realistic client is 2x2, cutting that again. The 46.1 Gbps figure requires 16 aggregated streams on one link, which essentially no shipping product does.

From 11.5 Gbps PHY to ~5 Gbps real

The gap between PHY and TCP throughput is MAC efficiency, and the honest budget for a good 4-stream 320 MHz deployment lands at 40–60% efficiency:

Applying ~45% efficiency to an 11.5 Gbps 4-stream radio gives ~5 Gbps — which is precisely the observed TCP throughput class of first-generation Wi-Fi 7 gear in good conditions, and matches the “12 Gbps PHY, 5 Gbps real” rule of thumb. The same math explains why 4096-QAM is a near-AP feature: 12-bit constellations need roughly 3 dB more SNR than 1024-QAM for the same error rate, so the highest rates only exist within one room of the AP, and the remaining benefit of the modulation is that clients sit longer at slightly lower-but-still-dense rates on the MCS curve.

Multi-Link Operation

One association, many radios

MLO changes the topology of a connection rather than any single radio. A client’s multi-link device (STA-MLD) associates once — one authentication, one keyspace, one IP address — with an AP-MLD that operates several affiliated APs on different bands or channels. The MAC layer then routes each aggregate or frame onto a link per packet, which yields three distinct wins: throughput aggregation across links, latency reduction because a frame is transmitted on whichever link is free instead of waiting for one contended medium, and reliability because a failed transmission on one link is retried elsewhere within the same connection. The aggregation datapath:

flowchart TD
    subgraph STA["Client STA-MLD: one MAC entity, two radios"]
        SCH["MAC link scheduler: per-packet link selection"]
        R1["Radio 1: 5 GHz, 160 MHz"]
        R2["Radio 2: 6 GHz, 160 MHz"]
    end
    subgraph AP["AP-MLD: one BSS, two affiliated APs"]
        L1["Link 1: 5 GHz, 160 MHz"]
        L2["Link 2: 6 GHz, 160 MHz"]
    end
    SCH --> R1
    SCH --> R2
    R1 -->|"frames of flow F"| L1
    R2 -->|"same flow F, in parallel"| L2
    L1 --> A1["Per-link block acks"]
    L2 --> A2["Per-link block acks"]

Link types: STR, non-STR, and asynchronous

Whether MLO aggregates or merely fails over is decided by the hardware’s ability to transmit and receive simultaneously without self-interference:

The distance between the client’s two radios and the antenna design decide which class a product falls into — self-interference from your own transmission on the other band is the physics problem STR has to beat. This is also why first-generation BE clients advertise MLO in specific combinations (a 2x2 + 2x2 dual-band design) rather than across all three bands. On the infrastructure side, the Linux mac80211 stack documents its multi-link device model, and support maturity varies by driver — a real constraint for anyone evaluating Wi-Fi 7 APs built on Linux: see the mac80211 documentation and the kernel wireless page.

Punctured Preamble

Wi-Fi 6’s answer to a partially occupied wide channel was to shrink to the largest contiguous clean width — a 320 MHz channel with one busy 20 MHz slice effectively became 160 MHz, halving capacity over 20 MHz of interference. 802.11be’s punctured preamble instead mutes the occupied 20 or 40 MHz subchannels and transmits the preamble over the full channel width with the muted parts flagged, so the remaining 280 (or 240…) MHz keeps carrying data. The receiver demodulates only the unpunctured resource units. This is “dynamic bandwidth” made real: it coexists with AFC-detected incumbents in 6 GHz, legacy 11ax neighbors, and radar-flagged segments without the cliff-edge fallback, and it composes with OFDMA and MLO (each link schedules around its own puncturing map). The cost is moderate — muted subchannels carry no data and some tone-plan irregularity — but it is far cheaper than the factor-of-two channel-width drop it replaces.

R-TWT and Bounded Latency

Target Wake Time in 802.11ax is a power-saving agreement: a client and AP agree on wake intervals, and the client sleeps between them. Restricted TWT (R-TWT) repurposes the mechanism for latency: the AP schedules restricted service periods in which latency-sensitive traffic gets protected access — other stations are asked to defer (and OFDMA scheduling enforces it), the member station transmits within the window, and the design goal is a bound on worst-case airtime delay rather than an average. For AR/VR, where the motion-to-photon budget is ~20 ms and the network’s share of it needs to be single-digit milliseconds at the 99th percentile, average-rate engineering does not help; bounded airtime does. R-TWT plus MLO is the intended pairing: if one link’s window is disturbed, the scheduler shifts traffic to the second link, and the puncturing mechanism keeps wide channels usable near incumbents — three 11be features aimed at the same variance problem.

Use Cases and the Client/AP Reality Check

The workload classes that justify Wi-Fi 7: untethered AR/VR headsets (high bitrate video per eye plus latency bounds); mesh backhaul, where MLO lets a node serve clients on one link while pushing backhaul on another simultaneously — the first Wi-Fi topology where backhaul and access do not time-share a radio; cloud gaming and 4K/8K streaming near the AP; and dense enterprise deployments, where OFDMA + R-U + R-TWT scheduling, not peak rate, is the actual product. In the mesh case the marketing claim “MLO backhaul” usually means the AP-MLD dedicates one band-pair to inter-AP links while the client-facing radios run separately — worth probing in interviews because it is a deployment decision, not a standard requirement.

The reality check that belongs in any Wi-Fi 7 pitch:

• 320 MHz is a 6 GHz, close-range feature. Two of the three channels exist in clean spectrum only in favorable regulatory domains, and phones rarely sustain it; most client traffic runs 80/160 MHz.
• 4096-QAM is a one-room modulation. The +3 dB SNR requirement confines peak rates to line-of-sight proximity; distance drops you down the MCS ladder quickly.
• MLO quality varies by product class. STR dual-radio clients deliver the latency win; NSTR designs mostly deliver failover. Ask which one a device implements before quoting “2x throughput”.
• First-wave devices are 2x2+2x2. The 46.1 Gbps figure needs 16 streams; real client PHY ceilings are ~2.9 Gbps per stream (320 MHz, 4096-QAM) and real TCP throughput a fraction of that.
• Linux AP stacks matured unevenly. MLO support in mac80211 and vendor drivers arrived progressively; verify driver status before specifying open-source Wi-Fi 7 infrastructure.

Interview Questions

1. Derive the 46.1 Gbps Wi-Fi 7 PHY rate and then explain why real throughput is ~5 Gbps. Start from 11ax: 9.6 Gbps at 160 MHz, 1024-QAM, 5/6, 8 streams = 1.2 Gbps per stream. Wi-Fi 7 doubles the channel (x2), raises 1024-QAM to 4096-QAM, i.e. 12 vs. 10 bits per symbol (x1.2), and doubles the streams to 16: 1.2 x 2 x 1.2 x 16 = 46.1 Gbps. A 4-stream radio tops at ~11.5 Gbps PHY, and MAC efficiency — contention, preambles, SIFS/ACK gaps, imperfect aggregation, retransmissions — keeps 40–60% of it, so ~5 Gbps TCP is the honest expectation, and a 2x2 client sees far less.

2. What does MLO change beyond throughput, and which link type delivers each benefit? MLO makes one association span multiple radios, so the scheduler can send on whichever link is free: that cuts latency (no waiting on one contended medium), improves reliability (per-packet failover within the same connection), and aggregates capacity. Only STR hardware — two radios with enough isolation to transmit and receive simultaneously — gets the aggregation and latency benefits in full; non-STR devices alternate between links and mostly realize the failover benefit, and asynchronous links degrade further toward simple redundancy.

3. What problem does the punctured preamble solve that Wi-Fi 6 could not? In 11ax, a busy or incumbent-occupied slice inside a wide channel forced a fallback to the largest contiguous clean width — one occupied 20 MHz block in a 320 MHz channel halved capacity to 160 MHz. 802.11be mutes the occupied 20/40 MHz subchannels, transmits the preamble over the full width with the punctured parts flagged, and uses the remainder — 280 MHz instead of 160. It matters most in 6 GHz, where AFC-detected incumbents and legacy neighbors make partially occupied wide channels the common case.

4. How is R-TWT different from 11ax TWT, and why does AR/VR need it? 11ax TWT is a power-saving agreement: a station sleeps between negotiated wake times. R-TWT schedules restricted service periods where latency- sensitive traffic gets protected airtime — other stations defer, and the AP’s OFDMA scheduling enforces the window — trading a little power and capacity for a bound on worst-case access delay. AR/VR has a ~20 ms motion-to-photon budget of which the network may take only single-digit milliseconds at the 99th percentile; average throughput optimizations do nothing for that tail, bounded airtime windows and MLO failover do.

5. Why does 4096-QAM need more SNR, and where is it actually useful? Each symbol carries 12 bits instead of 10, which moves decision points closer together in the constellation; holding the same error rate costs roughly 3 dB more SNR. That confines peak rates to short, line-of-sight links near the AP, so its practical value is shrinking the distance at which clients still run dense constellations — better real rates one and two rooms away — rather than a headline number that only exists in the same room as the AP.

6. A vendor advertises “Wi-Fi 7 mesh with MLO backhaul”. What should you verify? Which bands the backhaul MLO pair uses and whether it is STR (true simultaneous aggregation) or failover-grade NSTR; whether client-facing radios are separate from backhaul radios or time-shared; what the backhaul MCS actually sustains at your mounting distance (the 4096-QAM caveat applies to mesh links too); and regulatory limits on the 6 GHz segment in your region, since backhaul capacity plans built on 320 MHz channels assume spectrum you may only partially have.

Key Takeaways

• Wi-Fi 7 = 802.11be: 320 MHz channels (6 GHz only), 4096-QAM, 16 streams, MLO, punctured preamble, multiple RUs per user, 512-entry block ack, R-TWT — certified by the Wi-Fi Alliance as Wi-Fi 7.
• The 46.1 Gbps headline is 1.2 Gbps/stream x 2 (channel) x 1.2 (modulation) x 16 streams; real deployments live at 4 streams and 40–60% MAC efficiency, i.e. ~5 Gbps on a good day.
• MLO’s structural change is one association across multiple radios, with per-packet link selection buying latency and reliability — not just bandwidth.
• STR vs. non-STR is the hardware question that decides whether MLO aggregates or merely fails over; ask it before believing any MLO claim.
• Punctured preamble turns “one busy 20 MHz slice” from a halved channel into a muted subchannel — the enabler for 320 MHz in shared 6 GHz.
• R-TWT provides bounded, protected airtime for latency-critical traffic; it is the AR/VR story, paired with MLO failover.
• 6 GHz spectrum rules (AFC, LPI, regional splits) bound every deployment plan; 320 MHz channel counts are a regulatory question as much as an RF one.
• Client and driver maturity are the practical bottleneck: first-wave clients are 2x2+2x2 dual-band, and Linux mac80211 MLO support varies by driver.

References

1. IEEE 802.11 TGbe (Extreme High Throughput) task group status page — https://www.ieee802.org/11/Reports/tgbe_update.htm
2. IEEE 802.11 Working Group home (amendment index and timelines) — https://www.ieee802.org/11/
3. Wi-Fi Alliance — Wi-Fi 7 certification program (see “Discover Wi-Fi”) — https://www.wi-fi.org/
4. IEEE Std 802.11be, “Enhancements for Extreme High Throughput” (amendment text; access via IEEE Xplore)
5. Linux kernel mac80211 documentation (multi-link device model) — https://www.kernel.org/doc/html/latest/driver-api/80211/mac80211.html

Cross-References

• Wi-Fi — the generation overview table (11b through 11be), CSMA/CA, and OFDMA fundamentals this page builds on.
• 5G — the other shared-spectrum, latency-driven wireless story; useful contrast for 6 GHz sharing policy.
• Kernel Wireless Stack — cfg80211/mac80211 internals and where MLO support actually lands in Linux.
• Physical Layer — modulation, channel width, and SNR fundamentals behind the 4096-QAM and throughput arithmetic.