By Emil Olbrich
President, PrimeLime | VP Technology, SRG | Wireless Telecom Expert | 6G Academy Expert
Editor’s note: This article is adapted, with permission, from Emil Olbrich’s personal notes following 3GPP RAN1 #125, RAN2 #134, RAN3 #132, and RAN4 #119 in Dalian, China. The observations below reflect the author’s reading of the standards discussions and end-of-meeting chair notes, not official 3GPP records.
A few weeks ago I wrote that the Malta round was the moment 3GPP started acting like it learned something from 5G. Dalian is the follow-up, and it tells you a lot about whether that discipline holds when the decisions get expensive.
As a personal sidenote, Dalian, China was the first city I visited in China over 25 years ago to help run the CDMA2000 FOA (First Office Application) for Qualcomm with China Unicom. My how things have changed in technology and the city and country.
The short version: The physical layer is converging. The protocol stack is getting simpler in places that matter. The architecture study is now organized enough that you can see the shape of the fights. But the genuinely contentious calls — the ones with billions of dollars of deployment economics behind them — didn’t get made in Dalian. They got lined up for the June plenary in Singapore.
As always, a lot of what happened in Dalian was the usual grind: thousands of tdocs that ended “Noted,” feature-lead summaries that went nowhere, and twenty companies bringing the same idea in slightly different packaging. I’m skipping all of that. Here’s what’s worth your time.
A caveat before I start, same as last time: these are my personal observations from attending the meetings and from end-of-meeting chair notes. Several of the items below are “agreed unseen,” meaning the group accepted a direction but hasn’t reviewed the final text.
The Through-Line: Everything Points to June
If you read only one thing into this round, make it this. The 6G radio study has a hard interim milestone at the June plenary, and RAN3 has been told in plain language what it owes the TSGs by then: an interim result good enough for the plenary to decide the RAN-CN interface — point-to-point versus service-based — and the RAN internal split — CU-DU and CP-UP.
That is the whole ballgame for 6G network architecture, and it’s being decided in June. Not studied further…in theory this is the decision point for 6G.
Everything RAN3 did this round was about building the evaluation case each camp will carry into the plenary. And the “lean and streamlined” working principle — don’t over-option 6G the way 5G got over-optioned — is no longer an abstract pledge. It’s about to be tested against the first decision where picking one option means somebody’s preferred option loses. Watch whether the principle actually constrains the outcome or whether it becomes the thing everyone cites while voting for their own complexity.
I’ll come back to the architecture fight in Part 3, because it’s the headline. But the physical layer is where the most concrete progress happened, so let’s start there.
Part 1: The Physical Layer (RAN1 + RAN4)
The Lower-Tier Bandwidth Question Splits — Exactly the Way It Looked Like It Would
I’ve been tracking the smallest-maximum-UE-bandwidth decision for lower-tier 6G devices since Malta, because it matters a great deal to anyone building cost- and power-sensitive 6G endpoints — IoT, metering, wearables, the whole low-end. Malta produced an attempted down-select between 5 MHz and 20 MHz that failed to reach consensus.
Here’s what RAN1 actually agreed:
For TDD with 30 kHz SCS, RAN1 recommends 20 MHz RF and baseband bandwidth for both uplink and downlink. That’s settled (subject to the usual plenary endorsement). The TDD lower-tier device is a 20 MHz device. There’s an explicit note that this recommendation has no linkage to FDD — they fenced it deliberately.
For FDD with 15 kHz SCS, there is no down-select yet. RAN1 instead captured the reasoning behind three live alternatives:
- Alt 1 — 20 MHz RF and baseband for both UL and DL. Best system behavior when initial access has to fit inside a common bandwidth for all device tiers, but the largest device.
- Alt 2 — 5 MHz RF and baseband for both UL and DL. Roughly 8–10% additional modem-complexity reduction over Alt 1 for a single-FDD-band device, and 5 MHz is well understood for initial access, paging, and RACH from FDD deployment experience.
- Alt 3 — 20 MHz RF for both, 20 MHz baseband on DL, but only 5 MHz baseband on UL. The “asymmetric” option. The additional complexity reduction of Alt 2 over Alt 3 is no more than ~7%, and Alt 3 over Alt 1 is less than ~4%, for a single-band device.
The interesting part isn’t the percentages — it’s where RAN1 sent the decision next. Five separate sources observed that for multi-band devices, eliminating band-specific SAW filters in a SAW-less HD-FDD UL design can cut modem complexity by roughly 30–60%. That’s an order of magnitude more than the single-band savings, and it completely reframes the question. The narrow-UL alternatives only pay off in a big way if the device can go SAW-less, and whether that’s feasible is a RAN4 question.
So RAN1 said exactly that: the feasibility and benefits of SAW-less UE implementation for the different bandwidth alternatives is up to RAN4, and the RAN4 outcome may determine the RF and/or baseband bandwidth for HD-FDD UL. The HD-FDD uplink bandwidth is now formally in RAN4’s court.
And here’s the thing — RAN4 met the same week, in the same city, and discussed it. The device-type / smallest-max-CBW agenda had a full slate of contributions, including a Nordic Semiconductor paper specifically on SAW/BAW-less half-duplex design and a stack of operator and chipset views. No conclusion. So as of Dalian, the picture is:
- TDD lower-tier: 20 MHz, recommended.
- FDD HD-FDD UL: still open, with the narrow-UL options technically attractive but contingent on a SAW-less feasibility finding that RAN4 has started but not landed.
If you’re planning lower-tier 6G silicon, that’s your status. The complexity argument favors a narrow uplink baseband; the decision is gated on RAN4 proving the filter story holds. I’d expect this to keep moving over the August and October rounds rather than resolve cleanly in one shot.
Waveform: NR Under the Hood, With Two Things Worth Noticing
The waveform decision is mostly a “don’t reinvent it” outcome; CP-OFDM for the downlink and CP-OFDM plus DFT-s-OFDM for the uplink, as defined in 5G NR, are the 6G baseline. Enhancements to DFT-s-OFDM aren’t precluded, but the starting point is the NR waveform.
Two things break from a pure copy-paste, though, and both are worth flagging:
First, uplink DFT-s-OFDM gets multi-layer support, with a max rank of 2. In NR, the single-carrier uplink waveform is effectively single-layer. Giving DFT-s-OFDM two-layer MIMO with a real codebook — where each Tx port carries at most one layer — is a genuine capability add for coverage-limited uplinks that still want spatial multiplexing. The final codebook details get nailed down in the UL transmission agenda.
Second, 6G will support both semi-static and dynamic switching of the PUSCH waveform. Semi-static for all RRC states, dynamic for connected mode. That lets the network move a UE between CP-OFDM and DFT-s-OFDM based on conditions instead of pinning it. Details are still open, and it’s tied to the UE capability framework that hasn’t been decided yet, so don’t over-read it — but the intent is there.
There was also an agreement that frequency-domain spectrum shaping with transparent filtering is beneficial at least for π/2-BPSK, with an LS fired off to RAN4 to handle the filter requirements. This is a coverage play for the deep-uplink edge, and it’s the kind of thing where the PHY group sets the direction and RAN4 determines whether it’s actually buildable.
Channel Coding: LDPC Stays, But a New Base Graph Is on the Table for Multi-Gig
LDPC remains the data channel code for 6G. No surprise, and no Turbo-vs-LDPC re-litigation. What’s new is a study of a new base graph (informally BG3) to extend LDPC beyond the data-rate range NR was designed for. A broad coalition — Verizon, T-Mobile, CMCC, China Telecom, Samsung, ZTE, CATT, Apple, Qualcomm, MediaTek, Meta and others — is pushing the extension, and the reference data rates under study are eye-watering: on the order of 3 to 14 Gbps per component carrier (think 200–400 MHz, four layers, 1024QAM). It’ll work great in the lab with a single user…
This is the quiet signal that 6G peak rates are being designed for spectrum and antenna configurations that NR’s coding never anticipated. Whether BG3 actually gets standardized, or whether the network just configures BG1/BG2 per carrier, is still being argued — there are at least three “directions” on the table for when the new graph would even apply. But the fact that the data-rate ceiling is being raised, seems telling about what the high end of 6G is being dimensioned for.
Modulation: Uniform QAM Holds, Constellation Shaping Won’t Die, and FWA Keeps Showing Up
Looks like uniform QAM is the 6G baseline. Any geometric or probabilistic shaping, if supported at all, sits on top of uniform QAM as a network-configurable option — it doesn’t replace it.
But the shaping fight is still very much alive. A serious coalition — T-Mobile, CableLabs, Charter, DeepSig, InterDigital, Lenovo, MediaTek, Qualcomm, Samsung, Verizon — is keeping geometric shaping (1D/2D non-uniform constellations) and probabilistic shaping (ESS, CCDM, and friends) on the table. For someone like Deepsig, this is their bread and butter. RAN1 catalogued the options carefully and then noted, explicitly, that one of the live outcomes is “none of the enhancements are needed.” That option being written down is itself the lean principle showing its teeth: shaping has to justify its complexity against a uniform-QAM baseline that already works, and the group is refusing to assume it clears that bar.
And once again, uplink 1024-QAM keeps surfacing specifically for FWA. System-level evaluations from vivo and Ericsson show UL 1024-QAM getting selected something like 10–30% of the time in dense-urban and urban-macro, almost entirely for outdoor CPE, and overwhelmingly at rank 2. That seems like a real use case: the high-order-modulation push in 6G is being driven by fixed wireless economics, not by handsets.
PDCCH and the Control Channel: Built on NR, Then Streamlined
The downlink control channel work is a good example of the streamlining principle applied at the bit level. 6G PDCCH reuses NR’s structure as the benchmark almost everywhere — REG defined as one RB by one symbol, the length-31 Gold sequence for DMRS, NR’s CRC and payload scrambling procedures as the starting point, NR’s DMRS pattern as the evaluation benchmark. Where NR had a working answer, 6G is starting from it rather than from a blank page.
The new thinking is around making it more efficient: UE-transparent MU-MIMO on the PDCCH as the benchmark transmission scheme, a study of slot-level monitoring (restricting where in the slot a UE has to look), and explicit work on letting wideband and narrowband CORESETs coexist. None of this is glamorous, but 5G never got to do because it was always adding, never consolidating.
Native AI in the Air Interface
The AI/ML-for-air-interface work continues to mature on CSI and SRS. RAN1 keeps saying “strive for a unified design that works for both AI/ML and non-AI/ML schemes,” and “simplify the SRS carrier-switching framework compared to NR” by converging redundant triggering options.
Read that against the 5G experience. In 5G, AI/ML for the air interface arrived late and bolted on. In 6G, the group is explicitly trying to design one SRS framework that serves the model-based and the classical schemes together, and to use the occasion to throw out NR triggering complexity that nobody loved. Whether it holds is another matter — unification is easy to agree to in principle and hard to hold when every company’s model wants its own hooks.
Low-overhead SRS with network-side AI for CSI reconstruction is being studied, channel-property-information reporting (correlations, covariance, transform-domain statistics) is being defined as the substrate for AI-based CSI, and CSI-RS resource-sharing schemes are being worked to cut overhead. CJT calibration reporting from Rel-19 is being carried forward and extended. It does seem that the 6G CSI is being designed AI-aware from the start.
ISAC: UAV Detection Is the Flagship
Integrated sensing is still mostly in the evaluation-methodology phase, but the use-case priority is now obvious: detecting and tracking UAVs is the headline sensing application, with human, vehicle, and AGV tracking alongside it. The sensing modes are being pinned down — TRP mono-static, bi-static, UE mono-static — and the channel models and RCS assumptions for UAV targets are being agreed. T-Mobile/DT and AT&T/FirstNet are among the active contributors on scenarios, which tells you who sees the near-term value.
Separately, RAN1 agreed to identify at least one representative sensing-assisted-communication use case for study — sensing the environment to help beam management, CSI, mobility, or scheduling. That’s the more architecturally interesting direction long-term, because it folds sensing back into making the comms link better rather than treating radar as a bolt-on service. NTN-specific GNSS-less/resilient operation, by contrast, was held over with no contributions before the August round — worth noting for the satellite-resilience crowd that it’s parked, not dropped.
RAN4: The 6G Study Is Spun Up, But It’s Early
RAN4 #119 formally stood up its 6G study across system parameters, UE RF, spectrum, spectrum sharing, and RRM — channel raster, measurement gap design, unified measurements, UE-group RRM and the rest. But almost everything is still at the topic-summary and way-forward stage. The substantive RF numbers aren’t agreed yet. The most consequential RAN4 item for the broader narrative is the SAW-less device-type question I covered above, and that’s still open.
One cross-group item to flag: RAN1 sent RAN4 an LS on the phase-noise model for the 6G study. That’s the unglamorous plumbing that determines whether the higher-frequency 6G bands are actually usable, and it’s the kind of RAN1-RAN4 dependency that tends to quietly gate the whole upper-band story. Keep half an eye on it.
Part 2: The Protocol Stack (RAN2)
RAN2’s round was about the Layer 2 architecture and the connection-management model, and there’s a clear streamlining thread running through all of it.
Merged vs. Separate PDCP/RLC — The Header-Byte Fight
The big structural question for 6G L2 is whether to keep PDCP and RLC as separate layers (as in 5G) or merge them. RAN2 is studying both: the group wants a common protocol stack regardless of deployment option — split or monolithic, the UE’s stack shouldn’t change.
The technical core is the single-sequence-number idea. Collapsing to a single SN across the combined L2 header saves up to roughly two bytes per PDU and can simplify window management, at the cost of new complexity around segmentation, ARQ, and how the SN is carried across header boundaries. RAN2 captured the tradeoffs honestly — including that a single SN makes 6G-6G split-bearer dual connectivity harder, if the plenary even decides to keep split bearers. And critically, the L2 decision is entangled with the RAN3 CU-DU split decision.
A Single Paging Architecture
This is the cleanest example of 6G simplifying something 5G made complicated. RAN2 is converging on a single CN-triggered paging mechanism with a single paging area, with clustered paging (gaps between paging occasions within a cluster), the ability to address multiple UEs in one RRC paging message, and a stated requirement that 6G paging capacity be at least at 5G’s level. One mechanism, one area, instead of the layered paging arrangements that accumulated in 5G.
Faster State Transitions and RACH-less Mobility
A lot of RAN2 energy went into the inactive state and fast transitions. The direction: keep an inactive state with a resume-like fast transition based on stored UE context, study RACH-less resume using early timing advance and configured-grant resources, and reuse the early-TA and configured resources from cell-switch mobility to enable RACH-less handover. Network-triggered early UL synchronization is supported as baseline, with the source cell setting it up, the target cell computing the TA, and the UE doing a RACH-less mobility procedure with a valid TA in hand.
The honest open question RAN2 is sitting with is whether to even support data transmission in the inactive state (the 5G SDT story) or whether the right answer is just to make the transition to connected fast enough that you don’t need it.
Part 3: The Architecture (RAN3) — The Fight That Matters in June
RAN-CN Interface: P2P Got a Name, SBI Got “6G AMF,” and the Vote Is in June
The 6G RAN-CN point-to-point interface now has a name: Na (I wanted Nah or Nope 🙂 That got agreed with a broad co-sourcing list — Huawei, Nokia, Vodafone, Samsung, OPPO, Qualcomm, NEC, Deutsche Telekom, China Telecom, Jio, AT&T, Orange, FiberCop, NTT DoCoMo, Ofinno, China Unicom, Ericsson, Lenovo. Naming an interface is mundane, but a coalition that broad, agreeing on the P2P interface’s existence tells you the point-to-point camp is the default position heading into the vote.
The service-based-interface camp isn’t conceding, though. RAN3 is studying SBI for both “connectivity services” and “new services,” and for connectivity services the group is leaning toward calling the control-plane anchor a “6G AMF.” The SBI push is being carried by a recognizable cloud-native coalition — FiberCop, Jio, Qualcomm, Google, Charter and others — the same crowd that wants 6G to look more like a cloud-native service mesh and less like an evolved EPC.
Here’s the structurally important part: RAN3 explicitly stated it expects to make the final decision between P2P and SBI for connectivity services, and the plenary owes the TSGs that recommendation in June. They picked CMCC’s evaluation table as the baseline and started populating it across deployment, performance, and operational criteria — with several delegates pushing hard to keep “cloud friendliness” and “ease of upgrade” in the scoring, and others pushing to keep the evaluation grounded in first-order technical factors rather than turning it into a marketing exercise. That tension — cloud-native evaluation criteria versus traditional interface criteria — is the whole P2P-vs-SBI fight in miniature, and it gets resolved (or at least voted on) in June.
My read, and I’ll flag this as opinion rather than anything in the notes: the breadth of the P2P coalition and the “connectivity services first” framing both favor point-to-point as the connectivity-plane baseline, with SBI more likely to find its home in the “new services” plane. But the cloud-native camp has been patient and well-organized, and June is exactly the kind of plenary where a determined minority can extract a “study both” outcome that quietly defeats the lean principle. Watch the evaluation table — whoever controls the criteria controls the conclusion.
CU-DU Split: “Transparent to the UE” Turns Out to Mean Two Different Things
The other half of the June decision is the RAN internal split. RAN3 carried forward the agreement that a 6G higher-layer split, if standardized, shall be transparent to the UE — and then spent the meeting discovering that the room doesn’t agree on what “transparent” means. Two interpretations are on the table:
- Interpretation A: no air-interface (Uu) enhancements or impacts may be introduced specifically for the CU-DU split.
- Interpretation B: a unified UE implementation with no architecture awareness, but Uu enhancements are allowed provided they apply uniformly to split and monolithic deployments.
Ericsson and Nokia lean toward the stricter A (“the UE should have no knowledge whatsoever”), while Apple, Samsung, and others argue for B (a single UE procedure that works for both, which is functionally what “transparent” should mean). The compromise wording everyone could live with: it shall be agnostic to the UE whether it is connected to a split or non-split aNB. That papers over the A/B gap for now, but it’ll resurface the moment a specific Uu enhancement wants in.
On the substance, RAN3 agreed to study redistributing RRC functions between the CU and DU, on top of two sensible principles: a Central Anchor Principle (keep UE-context, security-key, capability, QoS, and inter-node coordination at the CU to preserve pooling gains) and an Intra-DU Principle (push the cell-local, real-time, latency-sensitive functions to the DU). Three redistribution options are being weighed, with several companies already moving to rule out the most aggressive “move everything including security to the DU” variants and there’s an open question about whether 5G’s CU-DU split is even a good baseline given where 6G mobility and MAC-layer security are heading. Detailed user-plane split work was deliberately postponed until the 6G UP protocol stack matures in RAN2 — which, again, points back to the L2 decision in Part 2. The whole thing is one interlocking dependency graph, and June is the node everything hangs off.
The aNB, the Xa Interface, and Agentic AI Creeping In
A few more architecture data points worth having. The 6G base station is the aNB (anodeB or rusty cell site), and the aNB-to-aNB interface is Xa — Xa-AP encoded in ASN.1 for connectivity services, Xa-U riding a point-to-point tunnel (pending CT4 input), with a load-management function for exchanging resource and traffic-load status between nodes. The Xa direction got agreed with another very broad co-sourcing list, which is the pattern this round: the existence and naming of the point-to-point interfaces is broadly settled, even as the RAN-CN P2P-vs-SBI question stays open.
On AI/ML, RAN3 is studying network energy saving and mobility optimization as the priority use cases, with L3-based handover assumed for 6G and AI applied to target-cell selection, handover timing, and trajectory/traffic prediction. The one to watch: a pCR on “Agentic AI” for the 6G network, co-sourced by ZTE, CMCC, Jio, Tejas, Pengcheng Lab, Seoul National University, FiberCop, China Unicom, China Telecom, TCL, Telecom Italia, and Ofinno. Agentic AI showing up in a RAN3 contribution list is a small thing today, but it’s the leading edge of the same conversation the core-network side has been having.
Next stops: the June plenary for the architecture call, then the working groups reconvene in Maastricht in late August and Jeju in mid-October. I’ll be there.
As always — this is my read of the standards work, not a substitute for the official records. I’ll see you in Singapore.
