R1-2602813
discussion
Maintenance on AI/ML for NR Air Interface
From Sharp
Summary
Sharp presents four technical proposals to resolve specification maintenance issues for AI/ML-based beam management in NR Rel-19, specifically focusing on RS-PAI reporting constraints, CSI computation timelines, and processing unit allocation. The document addresses gaps in TS 38.214 regarding PUCCH payload sizes for RS-PAI, the inclusion of AI/ML reports in Z/Z' timeline calculations, priority handling for CPU/APU resources, and the definition of computation time for model monitoring reports.
Position
Sharp requires that configurations resulting in 1 or 2-bit RS-PAI payloads be prevented for P-CSI and SP-CSI on PUCCH to avoid unspecified UE behavior, preferring network-side configuration restrictions over UE-side zero padding to alleviate UE burden. They propose that the determination of Zref and Z'ref timelines must consider the maximum computation time among both M and Mx updated CSI reports, ensuring AI/ML-based reports are not excluded from timing constraints. Sharp clarifies that for CSI reports with non-zero CPU and APU requirements, if the report is not considered within the available processing units, the UE is not required to update it, thereby preventing inconsistent priority handling where lower-priority reports are updated while higher-priority AI/ML reports are dropped. Finally, they propose reusing the legacy beam management CSI computation time for RS-PAI model monitoring reports to fill the current specification gap.
Key proposals
- Proposal 1 (RS-PAI Reporting on PUCCH): Adopt TP#1 to prevent configurations of nroftransmissionOccasion-r19 that result in 1 or 2-bit RS-PAI payloads for Periodic or Semi-Persistent CSI on PUCCH, as NR does not support PUCCH carrying 1 or 2 information bits other than HARQ-ACK/SR.
- Proposal 2 (Z and Z’ timeline for updated CSI reports): Adopt TP#2 to clarify that the maximum values Z and Z’ among both M (non-AI/ML) and Mx (AI/ML) updated CSI reports must be used to determine time locations Zref and Z'ref, ensuring sufficient processing time for AI/ML-based reports.
- Proposal 3 (Determination of updated CSI reports based on CPU and APU): Adopt TP#3 to clarify that if a CSI report requiring non-zero CPU and APU is not considered within the available processing units, it shall not be updated, preventing lower-priority reports from being updated while higher-priority AI/ML reports are skipped due to resource conflicts.
- Proposal 4 (CSI computation time of monitoring report): Adopt TP#4 to reuse the CSI computation time defined for legacy beam management for the model monitoring CSI report of RS-PAI, as no specific computation time is currently defined for this report type.