Content for TR 22.874 Word version: 18.2.0

0… 4 5… 5.2… 5.3… 5.4… 5.5… 6… 6.2… 6.3… 6.4… 6.5… 6.6… 6.7… 7… 7.2… 7.3… 7.4… 8… A… A.2 A.3 A.4 B C D…

8.1 KPI for AMMT services 8.2 Functional requirements for AMMT services
...

8 Consolidated potential requirements p. 61

8.1 KPI for AMMT services p. 61

In Table 8.1-1, Table 8.1-2 and Table 8.1-3, only the subset of experienced data rate values which are agreeable for the expected 5G enhancements are captured from the use cases.

The agreeable upper bound for split AI/ML inference is [1.1Gbit/s] in DL and UL respectively.
The agreeable upper bound for AI/ML model downloading is [1.1Gbit/s] in DL. The upper bound is [4Gbit/s] in DL if the downloading is only supported in hotspot coverage.
The agreeable upper bound for Federated Learning is [1.1Gbit/s] in DL and UL respectively.

Table 8.1-1: KPI Table of split AI/ML inference between UE and Network Server/Application function

Uplink KPI					Downlink KPI				Remarks
Max allowed UL end-to-end latency	Experienced data rate	Payload size	Communication service availability	Reliability	Max allowed DL end-to-end latency	Experienced data rate	Payload size	Reliability	Remarks
[CPR-001] 2ms	[CPR-002] 1.08Gbit/s (see note 1)	0.27 MByte	[CPR-003] 99.999 %	[CPR-037] 99.9%				[CPR-038] 99.999%	Split AI/ML image recognition
[CPR-004] [100ms]	[CPR-005] [1.5Mbit/s]				[CPR-006] [100ms]	[CPR-007] [150] Mbit/s	1.5 MByte /frame		Enhanced media recognition
		4.7Mbit/s			[CPR-008] 12ms	[CPR-009] 320Mbit/s	40kByte		Split control for robotics
NOTE 1: Only the values corresponding to AlexNet model is captured. NOTE 2: Communication service availability relates to the service interfaces, and reliability relates to a given system entity. One or more retransmissions of network layer packets may take place in order to satisfy the reliability requirement.

Table 8.1-2: KPI Table of AI/ML model downloading

Max allowed DL end-to-end latency	Experienced data rate (DL)	Model size	Communication service availability	Reliability	User density	# of downloaded AI/ML models	Remarks
[CPR-010] 1s	[CPR-011] 1.1Gbit/s	138MByte	[CPR-012] 99.999 %	[CPR-039] 99.9% for data transmission of model weight factors; 99.999% for data transmission of model topology			AI/ML model distribution for image recognition
[CPR-013] 1s	[CPR-014] 640Mbit/s	80MByte	[CPR-015] 99.999 %				AI/ML model distribution for speech recognition
[CPR-016] 1s	[CPR-017] 512Mbit/s / [4Gbit/s] (see note 1)	< 64MByte / 500MByte				Parallel download of up to 50 AI/ML models	Real time media editing with on-board AI inference
[CPR-018] 1s		536MByte			[CPR-019] up to 5000~ 10000/km2 in an urban area		AI model management as a Service
[CPR-020] [500ms]	[CPR-021] [100 Mbit/s]	[40MByte]	[CPR-022] 99.999 %				AI/ML based Automotive Networked Systems
[CPR-023] [1s]		[500]MByte					Shared AI/ML model monitoring
[CPR-024] 3s	[CPR-025] 450Mbit/s	[CPR-026] 170MByte					Media quality enhancement
NOTE 1: 512Mbit/s concerns AI/ML models having a size below 64 MB. 4Gbit/s concerns AI/ML models having a size below 500 MB where the model downloading is only supported in hotspot coverage. NOTE 2: Communication service availability relates to the service interfaces, and reliability relates to a given system entity. One or more retransmissions of network layer packets may take place in order to satisfy the reliability requirement.

Table 8.1-3: KPI Table of Federated Learning between UE and Network Server/Application function

Max allowed DL or UL end-to-end latency	DL experienced data rate	UL experienced data rate	DL packet size	UL packet size	Communication service availability	Remarks
[CPR-027] [1]s	[CPR-028] 1.0Gbit/s	[CPR-029] 1.0Gbit/s	132MByte	132MByte		Uncompressed Federated Learning for image recognition
[CPR-030] [1s]	[CPR-031] 80.88Mbit/s	[CPR-032] 80.88Mbit/s	10Mbyte	10Mbyte	[CPR-033] [99.9%]	Compressed Federated Learning for image/video processing
[CPR-034] [1s]	[CPR-035] [1.1Gbit/s]	[CPR-036] [500Mbit/s]	10MByte	10MByte		Data Transfer Disturbance in Multi-agent multi-device ML Operations

8.2 Functional requirements for AMMT services p. 63

Table 8.2-1: Functional requirments for AMMT services

CPR #	Potential Requirement	Original PR #	Comment
CPR 8.2-1	Based on operator policy, the 5G network shall provide the means to allow an authorized third-party to monitor the resource utilisation of the network service that is associated with the third-party. NOTE: Resource utilization in the preceding requirement refers to measurements relevant to the UE's performance such as the data throughput provided to the UE.	PR.5.5-001
CPR 8.2-2	Based on operator policy, the 5G system shall be able to provide an indication about a planned change of bitrate, latency, or reliability for a QoS flow to an authorized 3rd party so that the 3rd party AI/ML application is able to adjust the application layer behaviour if time allows. The indication shall provide the anticipated time and location of the change, as well as the target QoS parameters.	PR.5.5-002
CPR 8.2-3	Based on operator policy, 5G system shall be able to provide the means to predict (to the extent possible) and expose predicted network condition changes (i.e. bitrate, latency, reliability) per UE to the authorized third party.	PR.5.5-003
CPR 8.2-4	Subject to user consent, operator policy and regulatory constraints, the 5G system shall support a mechanism to expose monitoring and status information of an AI-ML session, (e.g. measured data rate/delay and other traffic analytics information), to a 3rd party AI/ML application. NOTE: such mechanism is needed for AI/ML application to determine an in-time transfer of AI/ML model	PR.6.4-001 PR.6.7-001	The two PRs are proposing to make 5GS provide information including measured data rate, delay, analytics result, and prediction for communication to 3rd party, for AI model downloading and training. It is proposed to merge the two PRs
CPR 8.2-5	5G system shall provide a means to supply event alerting to an authorized 3rd party, together with a predicted time of the event. (e.g., alerting about traffic congestion or UE moving into/out of a different geographical area). NOTE: A 3rd party AI/ML application may use the prediction information to minimize disturbance in the transfer of learning data and AI/ML model data.	PR.7.3-003	It is proposed to adopt it with the modification "learning agent → 3rd party" and rewording
CPR 8.2-6	The 5G system shall be able to support an authorised 3rd party to change aggregated QoS parameter values associated with a group of UEs, e.g. UEs of a FL group.	PR 7.4-001
CPR 8.2-7	Subject to user consent, operator policy and regulatory requirements, the 5G system shall be able to expose information to an authorized 3rd party to support the 3rd party to determine members of a group of UEs, e.g. UEs of a FL group, based upon criteria provided in the request from the 3rd party.	PR 7.4-001
CPR 8.2-8	The 5G system shall be able to expose aggregated QoS parameter values for a group of UEs to an authorized service provider.	PR.7.4-002 PR.7.4-003 PR.7.4-004	It is proposed to adopt the three potential requirements into CPRs
CPR 8.2-9	The 5G system shall be able to support collection of charging information for a group of UEs, e.g. UEs of a AI/ML FL group.	PR 7.4-004	The CPR is changed from the original PR to support the charging for FL group where the member may be dynamically changed.

9 Conclusion and recommendations p. 64

Regarding the Feasibility Study on traffic characteristics and performance requirements for AI/ML model transfer in 5GS, the TR analyses use cases of AMMT as follows:

Use cases on AI/ML operation splitting between AI/ML endpoints:
- Split AI/ML image recognition;
- Enhanced media recognition: Deep Learning Based Vision Applications;
- Media quality enhancement: Video streaming upgrade;
- Split control for robotics;
- Session-specific model transfer split computation operations.
Use cases on AI/ML model/data distribution and sharing over 5G system:
- AI/ML model distribution for image recognition;
- Real time media editing with on-board AI inference;
- AI/ML model distribution for speech recognition;
- AI model management as a Service;
- AI/ML based Automotive Networked Systems;
- Shared AI/ML model monitoring;
- Prediction of AI/ML model distribution.
Use cases on Distributed/Federated Learning over 5G system:
- Uncompressed Federated Learning for image recognition;
- Compressed Federated Learning for image/video processing;
- Data Transfer Disturbance in Multi-agent multi-device ML Operations;
- Group Performance "Flocking" Use Case.

It is recommended to proceed with normative work, and include the potential new requirements identified by this TR into a new TS. The consolidated potential requirements in Clause 8 are candidates for the normative requirements.