R1-2600151
discussion
Aspects of downlink-based CSI acquisition
From Huawei
Summary
This 3GPP RAN1 contribution from Huawei presents 23 Proposals and 16 Observations on downlink-based CSI acquisition for the 6G Radio (6GR) study, addressing challenges like high overhead and complexity in massive MIMO systems, especially at around 7 GHz. The document proposes a new framework centered on leveraging stable long-term channel information (PAS, PDP) to enable sparse CSI-RS, explicit feedback, unified codebooks, cell-specific precoded CSI-RS, and AI/ML-based CSI prediction with continuous learning.
Position
Huawei proposes a fundamental shift from the 5G NR instantaneous snapshot-based CSI acquisition to a 6GR framework that separates and exploits long-term channel information (PAS, PDP) as prior knowledge. They present technical case for using long-term channel information to enable sparse CSI-RS designs (e.g., 1RE/16RB/1port density) while maintaining or improving channel estimation accuracy, claiming up to 78% SGCS improvement and 120% MU-MIMO SE gain over legacy methods. They require studying a unified codebook structure based on a generic W = W1 * W * Wf architecture that supports both non-precoded and precoded CSI-RS across single/multi-TRP scenarios, and propose explicit CSI feedback as a primary mechanism to reduce UE complexity by shifting SVD computation to the base station. They propose studying cell-specific precoded CSI-RS where each port is transmitted by all TXRUs with a cell-specific beamforming vector to improve per-port SINR and enable flexible resource sharing among UEs with different measurement capabilities. For AI/ML, they prioritize studying frequency/spatial domain CSI prediction with sparse CSI-RS (sub-case A) and strongly propose incorporating long-term channel information as an AI/ML input to resolve 'one input to multiple outputs' ambiguity, claiming this can reduce model size by 75% and training data by 55%. They argue that two-sided AI/ML models for CSI compression should be discussed with lower priority in 6GR due to inter-vendor collaboration risks and duplication with NR Rel-20 work. They oppose moving CSI from L1 to L2 uplink control signaling, citing efficiency, reliability, latency, and coding performance advantages of L1 UCI transmission with Polar/RM codes.
Key proposals
- Proposal 1 (Sec 2.1.3): 6GR shall study and address the following challenges for DL CSI acquisition: reliance on instantaneous measurement and report; degradation of channel measurement accuracy due to large path loss around 7GHz; managing the increase in RS overhead, CSI feedback overhead, and UE complexity; CSI resource sharing among UEs with different measurement capabilities.
- Proposal 2 (Sec 3.1.2): 6GR shall study the acquisition and utilization of long-term channel information to facilitate MIMO operations.
- Proposal 5 (Sec 3.4): Cell-specific precoded CSI-RS transmission (e.g., each port is mapped to/transmitted by all the TXRUs with a cell-specific beamforming vector) should be studied in 6GR.
- Proposal 6 (Sec 4.1): 6GR should study CSI-RS design supporting up to 512 ports per resource.
- Proposal 8 (Sec 4.2): 6GR MIMO shall study optimizing CSI-RS overhead and performance trade-off via using long-term channel information.
- Proposal 11 (Sec 4.3.2): For frequency and/or spatial domain CSI prediction with sparse/low overhead CSI-RS with AI/ML, study long-term channel information as assistance information of AI/ML input.
- Proposal 12 (Sec 4.3.4): For frequency and/or spatial domain CSI prediction with sparse/low overhead CSI-RS with AI/ML, study the performance and benefits of continuous learning.
- Proposal 14 (Sec 5.1): 6GR MIMO shall study the feedback and utilization of long-term channel information for efficient CSI measurement and feedback.
- Proposal 15 (Sec 5.2.2): 6GR MIMO shall study explicit CSI feedback for high-precision CSI acquisition, with assistant of long-term channel information-based.
- Proposal 17 (Sec 5.3): 6GR should study DMRS-based CSI acquisition to improve inter-cell interference measurement.
- Proposal 19 (Sec 5.4.2): For CSI compression and feedback with AI/ML, the two-sided model solutions are discussed with lower priority.
- Proposal 20 (Sec 5.4.3): 6GR should support CSI as L1 UL control information.
- Proposal 22 (Sec 6.1): For evaluation of enhancements on downlink based CSI acquisition: Both SLS and LLS are considered. Use Table 8 and Table 7 as starting point. Consider intermediate KPIs, such as SGCS/NMSE; link level KPs, such as BLER/SE/throughput; and system level KPIs, such as cell average and edge downlink SE, cell average and edge UPT.
- Proposal 23 (Sec 6.3): Consider the following aspects of AI/ML specific evaluation methodology for 6GR CSI prediction: Link level simulation and system level simulation can be considered; For complexity KPI, at least consider FLOPs/MACs, number of parameters; Generalization and scalability performance, including the aspects of deployment scenario, UE speed, indoor/outdoor distribution, antenna configurations, sparse CSI-RS pattern (frequency and/or spatial domain) and density, etc.; Imperfect/non-ideal factors at least including noising and channel estimation error for model input and label.
Revision chain
1 versions in this meeting · oldest first
-
R1-2600151 ← you are here discussion revised