3GPP Release 18: 5G-Advanced Phase 1 ā Intelligence and Efficiency
Release 18 marks the beginning of the 5G-Advanced era ā 3GPP's branding for the next phase of 5G evolution, analogous to how LTE-Advanced extended LTE. The defining theme is intelligence: AI and machine learning are embedded directly into the NR air interface specifications for the first time, not as a network management overlay but as standardised Layer 1 and Layer 2 procedures exchanged between the gNB and UE.
Overview
The 5G-Advanced branding signals more than a marketing pivot. 3GPP uses it to indicate that the standard has reached a level of maturity where incremental radio enhancements give way to architectural intelligence ā the network learns from its environment rather than relying entirely on human-designed algorithms. Rel-18 establishes the framework, interfaces, and first three AI/ML use cases that will be developed and refined through Rel-19 and Rel-20.
Alongside AI/ML, two major operator priorities shaped Rel-18: XR traffic optimisation (AR and VR headsets create scheduling challenges that standard NR handles poorly) and network energy efficiency (electricity cost is the primary operating expense for mobile networks and regulators are setting mandatory efficiency targets). Both received dedicated work items with significant air-interface changes.
AI/ML in the Air Interface
Rel-18 standardises three specific AI/ML use cases within the NR protocol stack. The defining characteristic is that these are standardised interfaces: a UE from vendor A and a gNB from vendor B must agree on how AI/ML models are represented, exchanged, and evaluated ā the same interoperability requirement that applies to every other NR feature.
Channel State Information describes the radio channel between gNB and UE ā in a massive MIMO system with 64 antenna ports, the full CSI matrix is large and expensive to report over the air. The UE runs a neural network encoderthat compresses the CSI matrix before transmission; the gNB runs the paired decoder to reconstruct it. The same channel accuracy is preserved with 4ā8 times less uplink overhead, freeing resources for data. The encoder and decoder are trained as a matched pair and the gNB provides the UE with its decoder's expected input format via standardised signalling.
In standard NR, beam management requires periodic sweeping (P1/P2/P3) to identify the best beam ā a process that takes measurement time and consumes reference signal resources. An ML model at the gNB or UE can predict which beam will be optimal in the next slot based on the history of received signal measurements, without waiting for a full sweep cycle. At mmWave, where beams are narrow and the best beam changes rapidly with small movements, ML prediction reduces beam failure rate and measurement overhead simultaneously.
Traditional 5G positioning algorithms (TDOA, AoA, AoD) use geometric calculations that degrade in environments with dense multipath ā indoor settings, tunnels, and dense urban canyons where signals bounce off multiple surfaces before arriving. An ML model trained on the specific environment maps the received signal features (RSRP, RSRQ, timing measurements from multiple TRPs) directly to a position estimate, outperforming geometric algorithms in these challenging scenarios.
Standardising the use case interfaces is only half the problem. Rel-18 also defines how models are managed over their lifecycle: how a model is transferred to the UE over-the-air via dedicated NR signalling channels; how the network monitors inference quality (detecting when a model's performance has drifted from its training conditions); and how the network triggers a model update or replacement when quality falls below a threshold.
XR Traffic Optimisation
Extended Reality headsets ā AR glasses and VR headsets ā generate a traffic pattern that standard NR schedulers were not designed for. A rendering frame at 90 fps arrives every 11 ms as a burst of 5ā20 packets that must all be decoded and rendered before the next frame deadline. If even one packet in a frame misses its deadline, the entire frame is dropped by the codec ā causing a visible stutter that triggers motion sickness in VR users. Standard NR schedulers optimise for throughput, not per-frame deadline compliance.
Rel-18 introduces PDU Set awareness to address this. A PDU Set is the group of IP packets that collectively form one video frame. The network and UE treat the PDU Set as a scheduling unit with an associated deadline ā not as a collection of independent packets. Three specific enhancements:
- PDU Set identification and marking: the application layer marks packets belonging to the same frame with a common PDU Set ID. The gNB scheduler can see these marks and group the packets accordingly.
- Deadline-aware scheduling: the scheduler prioritises delivering all packets in a PDU Set before the frame deadline, even at the cost of delaying other lower-priority traffic. Transmissions are timed to arrive just before the deadline rather than spread uniformly across the frame interval.
- XR-aware power saving: since XR rendering is frame-rate-locked, the UE's radio can sleep predictably between frame bursts. The gNB is informed of the UE's frame rhythm and does not schedule transmissions or expect uplink during the sleep windows, extending headset battery life significantly.
Network Energy Efficiency
A large mobile operator's network can consume over ā¬100 million per year in electricity, with the radio access network responsible for roughly 70ā80% of that consumption. Rel-18 defines structured gNB sleep modes that enable substantial power reduction during low-traffic periods without breaking connected UEs. Three levels of sleep granularity:
In all three modes, the gNB must maintain or explicitly suspend mandatory reference signals (SSB, CSI-RS). L3 sleep, which provides the greatest energy saving, requires the most careful coordination ā a UE that doesn't receive the sleep notification would declare radio link failure and re-attach, negating the energy saving and causing a service interruption. The Rel-18 specification defines the exact signalling that prevents this. Energy saving potential across off-peak hours: operators report 30ā50% reduction in RAN power consumption.
Coverage Enhancements
5G NR's higher frequency deployments ā 3.5 GHz, 26 GHz ā have inherently worse propagation than 4G LTE at 800 MHz or 2.1 GHz. A rural area covered by a single LTE 800 MHz macro might require four or five 3.5 GHz NR macros to achieve equivalent coverage. Rel-18 improves the link budget through several physical-layer enhancements:
- PUSCH repetition with incremental redundancy: the UE transmits the same PUSCH block across multiple consecutive slots. The gNB soft-combines the copies using incremental redundancy ā each retransmission carries a different subset of the coded bits, so the combined information grows with each copy. This improves uplink coverage by 6ā10 dB without requiring higher transmit power from the UE, extending range without shortening battery life.
- DMRS bundling across slots: linking pilot symbols (DMRS) across multiple slots allows the gNB to perform more accurate channel estimation in low SNR conditions. Better channel estimation directly improves the ability to use higher modulation orders at the cell edge, increasing coverage-limited throughput.
- Flexible minimum allocation size: reducing the minimum number of symbols in a PUSCH/PDSCH allocation allows more flexible scheduling of coverage-limited UEs in resource blocks where other UEs occupy most symbols. Coverage-limited UEs get more efficient use of available gaps.
eRedCap ā Even Simpler IoT
Enhanced Reduced Capability (eRedCap) takes the Rel-17 RedCap profile and reduces it further for the very-low-cost sensor market. The key change: downlink bandwidth is reduced from 20 MHz to 5 MHz. This targets IoT applications that need only kilobits to low megabits per second ā environmental monitoring nodes, building management sensors, utility metering devices.
A 5 MHz RF frontend is dramatically cheaper than a 20 MHz one: smaller filter, simpler ADC, lower power consumption in active mode. The target chipset cost is competitive with LTE-M at sub-$2, while eRedCap devices retain all mandatory 5G NR features: network slicing support (connecting to the right slice for the industrial application), NR security (5G AKA authentication and 256-bit ciphering), and 5GC connectivity (full PDU session management through the SMF). LTE-M offers none of these at equivalent cost.
Why Rel-18 Mattered
- AI/ML standardisation in the air interface means all vendors must support the same model interfaces ā enabling ecosystem development where a UE from one chipset vendor works with a gNB from another using AI/ML-enhanced CSI feedback or beam prediction. Without standardisation, each vendor pair would require proprietary alignment.
- XR optimisation is commercially critical as AR/VR headsets move from niche to mainstream consumer electronics ā Apple Vision Pro, Meta Quest, and enterprise XR devices all generate the frame-burst traffic pattern that Rel-18's PDU Set scheduling addresses. Without it, XR over 5G produces poor user experience even at high measured throughput.
- gNB sleep modes give operators a tangible path to meeting EU energy reduction targets ā European operators are committed to 50% energy reduction per bit by 2030. L2 and L3 sleep modes, applied at scale across night-time low-traffic periods, represent the largest single available efficiency lever.
- eRedCap extends 5G's addressable IoT market to compete with LTE-M's cost profile while retaining 5G's superior quality-of-service framework ā enabling operators to serve both consumer and industrial IoT from a single 5G NR network without maintaining a parallel LTE-M carrier.