AI/ML for NR Air Interface

agreement

The following TP for TS 38.212 Clause 6.3.1.1.2 is endorsed. Reason for change: For P/SP CSI reporting on PUCCH for performance monitoring, it is possible that the number of information bits of RS-PAI becomes less than 3 bits. However, NR specifications do not support PUCCH carrying 1 or 2 information bits other than HARQ-ACK/SR. Summary of change: For P/SP CSI reporting on PUCCH for performance monitoring, the number of information bits of related CSI is zero padded to 3 bits. Consequences if not approved: For P/SP CSI reporting on PUCCH for performance monitoring, 1-bit or 2-bit payload size of RS-PAI is not supported. The corresponding final CR in: (Rel-19, 38.212, NR_AIML_air-Core, CR0246, Cat F) is agreed.

agreement

The following TP for TS 38.214 Clause 5.2.1.6 is endorsed. Reason for change: CPU duration for beam management related UE-side data collection is not captured correctly. In current specification, periodic CSI report for UE-side data collection for beam management is not supported. Summary of change: Reflect the RAN1 agreements accurately in Section 5.2.1.6 of TS 38.214. Include periodic CSI report for UE-side data collection for beam management. Consequence if not approved: Incorrect UE behaviour for CPU duration for UE-side data collection for training. The corresponding final CR in: (Rel-19, 38.214, NR_AIML_air-Core, CR0751, Cat F) is agreed.

conclusion

There is no RAN1 consensus on the following: For virtual resource blocks mapping of PDSCH in TS38.211, rate matching of the resource elements in the corresponding physical resource blocks are not performed around the NZP CSI-RS only configured in Set A for CSI report for inference. Final summary in R1-2603261.

agreement

For the standardized SQ codebook when NW does not exchange a codebook, support [1/8, 3/8, 5/8, 7/8] for Q=2 and [1/4, 3/4] for Q=1.

agreement

For precoding matrix feedback via two-sided model, support upto 128 ports, upto rank 4 and upto 19 subbands.

agreement

Support two-part UCI for CSI feedback via two-sided model, where UCI part 1 contains CQI, RI and CRI. UCI part 2 contains the PMI (i.e., CSI feedback of all layers). Layer 1 to are mapped to Group 1, Layer to layer are mapped to Group 2. Note: The bits sequence is mapped based on their indices inside each layer, followed by the bit sequence of the next layer. The layers are ordered from the strongest to the weakest.

agreement

For NW side data collection, support target CSI quantization Down-selection between 1a (real-imaginary) and 1b (amplitude-phase), i.e., {1a-33, 1b-34}. Presented in Thursday session. Potential agreement from session notes in R1-2603273: For performance monitoring: Support using paired CSI feedback and target CSI samples for performance monitoring Support using L3 signaling framework to collect at least target CSI samples. Detailed signaling upto RAN2 FFS how to associate target CSI sample with CSI feedback sample. Support using L1 signalling FFS: Details of L1 signalling UE assisted monitoring Target CSI reporting via L1 signaling with legacy PMI codebook Target CSI reporting using L1 signaling with high resolution quantization of partial target CSI Strong concern from Nokia and CATT Nokia and CATT state that Network sides performance monitoring can be supported by implementation means which makes this proposal or any specification impact is not needed. Other companies claim based on WID, we have to specify performance monitoring. CSI feedback enhancement, encompassing (two-sided model) [RAN1, RAN2]: CSI spatial/frequency compression without temporal aspects (“Case 0”), Specify necessary signalling/mechanism(s) as per the identified potential specification impacts in FS_NR_AIML_Air , including: Model pairing procedure including ID and applicability reporting Inference aspects including target CSI type, measurement and report configuration, CQI RI determination, payload determination, quantization configuration codebook, UCI mapping, CSI processing criteria and timeline, priority rules for CSI reports NW and UE side data collection for training, Target CSI format Note The framework defined in BM and CSI prediction use cases could be reused Specify performance monitoring

agreement

For performance monitoring: Support using paired CSI feedback and target CSI samples for performance monitoring, and further investigate the feasibility of following options and downselect one, Option 1: Support using L3 signaling framework to collect target CSI samples. Companies to report details about how to associate target CSI sample with CSI feedback sample. Option 2: Support using L1 signalling, considering UE assisted monitoring CSI reporting via L1 signaling with legacy PMI codebook CSI reporting using L1 signaling with high resolution quantization of partial target CSI

agreement

For Option 4-1 under Direction A in AI/ML based CSI compression, when the target CSI is precoding matrix with port-subband domain representation and the target CSI set k is associated with set of CSI feedback set, Support per-layer CSI feedback payload parameter pair For layer 1/2/3, support {32, 2}, {64, 2}, {96,2}, {128, 2}, {192, 2}, {32,1} For layer 4, support {32, 2}, {64, 2}, {96,2}, {32,1} FFS: The number of layers in the CSI feedback set is <= the number of layers in the target CSI set. Note: NW may only exchange a subset of the CSI feedback payload from the list.

agreement

For Option 4-1 under Direction A in AI/ML based CSI compression, the performance targets are exchanged per layer per CSI feedback set (), where the CSI feedback set () is the mth out of Mk CSI feedbacks set corresponding to the kth target CSI set. Presented in Wednesday session.

agreement

For Option 4-1 under Direction A in AI/ML based CSI compression, One-to-one mapping between pairing ID and exchanged dataset. There is no linkage between different data sets. Final summary in R1-2602693.

agreement

Prioritize the following cases in the RAN1 study on UL coverage improvements through low UL PAPR enhancement for DFT-s-OFDM, based on Table in 5.3.1 of R1-2603312: FDSS [for π/2 BPSK, QPSK, 16QAM] Opt1: Transparent filter Filter defined through transmitter requirement such as spectrum flatness by RAN4 Opt2: non-transparent filter Filter coefficients specified in RAN1 Note: This includes all the listed FDSS filtering variants FDSS for [π/2 BPSK, QPSK, 16QAM] with spectral extension For FDSS part, Opt1: Transparent filter Filter defined through transmitter requirement such as spectrum flatness by RAN4 Opt2: non-transparent filter Filter coefficients specified in RAN1 Note: This includes all the listed variants of the spectrum extension FDSS for π/2 BPSK with spectral truncation For FDSS part, Opt1: Transparent filter Filter defined through transmitter requirement such as spectrum flatness by RAN4 Opt2: non-transparent filter Filter coefficients specified in RAN1 Note: This includes all the listed variants of the spectrum truncation Note: At least one variant of spectral extension, offset-QPSK with smaller DFT size can result in the same signal as the FDSS for π/2 BPSK-ST Note: companies are encouraged to report whether transparent CFR is used or not

conclusion

The study on the following listed proposals is noted. Further discussion on these is not planned under the waveform agenda item in RAN1 unless tasked by RAN at or after RAN#112. Observations (with references added on Friday): Max-rank = 2 for UL DFT-s-OFDM Companies report negligible SNR loss from DFT-s-OFDM with larger number of aNB Rx antenna elements, such as at least 16 or 32. Link level loss with 4 layers has not been studied. A lower number of aNB antenna elements may result with SNR loss at the receiver All companies acknowledge that when the UE is Tx power limited, there is a Tx power gain from DFT-s-OFDM, highest reported gain of around 2.5dB. The system level impact depends on scenario and scheduling strategy due to vastly differing probabilities for UE to be power limited and use rank=2, as well as the used traffic model and the assumed number of aNB Rx antenna elements. The reference setup to compare with varied and had also an impact to the findings, including also system loading, link adaptation, scheduling strategy, the used UE antenna model and the number of aNB Rx antenna elements. Two companies reported cases with average system level throughput loss [R1-2601827, R1-2602177], of which one company indicated cell edge throughput loss [R1-2602177]. One company reported no average and cell-edge system level throughput gain [R1-2602254], while 6 companies reported cases with average system level throughput gains and/or cell edge throughput gains [R1-2602004, R1-2603056, R1-2602468, R1-2603008, R1-2603286 , R1-2601921]. Codebook for multi-layer UL DFT-s-OFDM The lower PAPR property of DFT-s-OFDM is diminished with coherent precoders due to mixing of two (or more) layers for each power amplifier. RAN1#124 agreed to focus the DFT-s-OFDM studies to non-coherent CB only (with the exception of lower number of layers than Tx antennas) for 2Tx UE for waveform evaluation purposes. Four companies report moderate system level gains for DFT-s-OFDM non-coherent codebook only relative to CP-OFDM non-coherent codebook [R1-2603056, R1-2602468, R1-2603008, R1-2603286], while two companies report system level loss for DFT-s-OFDM non-coherent codebook only relative to CP-OFDM non-coherent codebook [R1-2601827, R1-2602177]. One company reports system level gains also for DFT-s-OFDM with coherent codebook relative to CP-OFDM with coherent codebook [R1-2602468]. Six companies report system level gains for DFT-s-OFDM with non-coherent multi-layer codebook and coherent single-layer codebook relative to CP-OFDM with coherent single/multi-layer codebook [R1-2602426, R1-2602468, R1-2602004, R1-2603008, R1-2603286, R1-2601921], while one company reports no system level gain for DFT-s-OFDM with non-coherent multi-layer codebook and coherent single-layer codebook relative to CP-OFDM with coherent single/multi-layer codebook [R1-2602254]. Dynamic waveform switching One company observes comparable system level performance for DWS among DFT-s-OFDM for both 1- and 2-layer and CP-OFDM with both 1-layer and 2-layer compared to DWS among DFT-s-OFDM for only 1-layer and CP-OFDM for both 1- and 2-layer for full buffer traffic only and observes gain for non-full buffer traffic [R1-2603286]. One company observes system level loss of DFT-s-OFDM for both 1- and 2-layer compared to DFT-s-OFDM for only 1-layer and CP-OFDM for both 1- and 2-layer with waveform switching [R1-2601827]. One company observes comparable system level performance for DWS among DFT-s-OFDM for both 1- and 2-layer and CP-OFDM with both 1-layer and 2-layer compared to DWS among DFT-s-OFDM for only 1-layer and CP-OFDM for both 1- and 2-layer. [R1-2602254]. One company observes system level performance loss of DFT-s-OFDM for 1-layer and DFT-s-OFDM/CP-OFDM for 2-layer compared to DFT-s-OFDM for only 1-layer and CP-OFDM for both 1-layer and 2-layer with waveform switching [R1-2602177]. Note: this observation is not intended to be captured as it is in TR. Presented in Friday session

New contributors this meeting

Company positions

Open issues