R1-2410152
discussion
AI/ML based CSI Compression
From Google
Summary
Google's technical document on ML-based CSI compression for 5G NR presents 15 proposals covering CSI report content, processing units, model monitoring, data collection, and inter-vendor collaboration. The document addresses key aspects of AI/ML integration into NR air interface including compressed W2 reporting, dual processing units for ML inference, and standardized transformer-based reference models.
Position
Google advocates for a comprehensive AI/ML-based CSI compression framework that prioritizes practical deployment considerations. They push FOR: (1) transformer-based standardized models over other architectures, (2) flexible hybrid AI/ML and non-AI/ML approaches based on rank indicators, (3) separate processing units for ML inference vs channel estimation, and (4) dataset sharing (target CSI + CSI feedback) for inter-vendor collaboration. They push AGAINST: requiring common encoders across UEs and using SCS as a performance monitoring metric, instead favoring hypothetical BLER.
Key proposals
- Proposal 1 (CSI Report content): For case 2 and case 3, support the UE to report W1 and compressed W2 for a configured rank, with compressed W2 calculated based on AI/ML
- Proposal 2 (Priority): The priority for non-ML based CSI report should be higher than the priority of ML based CSI report
- Proposal 3 (CSI Processing Unit): Support CPU occupancy rule for ML based CSI based on two types processing unit - Type1 CPU (MPU) for channel estimation and Type2 CPU (IPU) for ML inference
- Proposal 4 (AI/ML model monitoring): Introduce new report quantity PMI only and support subband L1-SINR reporting, with CSI state indication for invalid measurements in low SINR cases
- Proposal 5 (SRS monitoring): Support SRS linked with CSI-RS report configuration for ML based CSI and burst based SRS with frequency hopping for coverage-limited UE
- Proposal 6 (UE-side monitoring): Introduce hypothetical BLER as metric for performance calculation with configuration of precoded CSI-RS
- Proposal 7 (NW data collection): Support configurable number of layers for report for NW side data collection for performance monitoring
- Proposal 8 (Ground truth): Support reporting singular values for the ground-truth CSI
- Proposal 11 (UE data collection): Support maintaining same understanding between NW and UE on measurement timing via NW configuration or UE request options
- Proposal 12 (Hybrid approach): Support hybrid AI/ML and non-AI/ML CSI where UE reports AI/ML based CSI for small RI and Type1 codebook for large RI
- Proposal 13 (Inter-vendor Direction A): Prioritize option 4-1 with dataset containing target CSI and CSI feedback
- Proposal 14 (Inter-vendor Direction B): Should not require common encoder across UEs for NW side encoder parameter sharing
- Proposal 15 (Inter-vendor Direction C): Fully standardized reference model should be based on transformer architecture only