Tech-invite  3GPPspace  IETFspace
21  22  23  24  25  26  27  28  29  31  32  33  34  35  36  37  38  4‑5x

Content for  TS 22.261  Word version:  19.6.0

Top   Top   Up   Prev   Next
0…   4…   6…   6.4…   6.8…   6.12…   6.15…   6.16…   6.22…   6.26…   6.30…   6.37…   6.41…   6.43…   6.46…   7…   7.4…   7.9   7.10…   7.11…   8…   D…   D.3…   F…   G…   H   I   J… 
7.10 KPIs for AI/ML model transfer in 5GS  7.10.1 KPI requirement for direct network connection  7.10.2 KPI requirement for direct device connection 
...

 

7.10  KPIs for AI/ML model transfer in 5GS |R18|  p. 104

7.10.1  KPI requirement for direct network connection |R19|  p. 104

The 5G system shall support split AI/ML inference between UE and Network Server/Application function with performance requirements as given in Table 7.10.1-1.
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
15 ms144 Mbit/s0.27 MByte99.999 %99.9 %99.999 %Split AI/ML image reco­gni­tion
100 ms1.5 Mbit/s100 ms150 Mbit/s1.5 MByte /frameEn­hanced media reco­gni­tion
4.7 Mbit/s12 ms320 Mbit/s40 kByteSplit control for robo­tics
NOTE 1:
Communication service availability relates to the service interfaces, and reliability relates to a given system entity. One or more retransmissions of network layer packets can take place in order to satisfy the reliability requirement.
The 5G system shall support AI/ML model downloading with performance requirements as given in Table 7.10.1-2.
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
1s1.1 Gbit/s138 MByte99.999 %99.9% for data trans­mission of model weight factors; 99.999% for data trans­mission of model topo­logyAI/ML model distri­bution for image reco­gni­tion
1s640 Mbit/s80 MByte99.999 %AI/ML model distri­bution for speech rec­ogni­tion
1s512 Mbit/s (see note 1)64 MByteParallel down­load of up to 50 AI/ML modelsReal time media editing with on-board AI infe­rence
1s536 MByteup to 5000~ 10000/km² in an urban areaAI model mana­gement as a Service
1s22 Mbit/s2.4 MByte99.999 %AI/ML based Auto­motive Networ­ked Systems
1s500 MByteShared AI/ML model moni­toring
3s450 Mbit/s170 MByteMedia qua­li­ty enhan­cement
NOTE 1:
512Mbit/s concerns AI/ML models having a payload size below 64 MB. TBD for larger payload sizes.
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 can take place in order to satisfy the reliability requirement.
The 5G system shall support Federated Learning between UE and Network Server/Application function with performance requirements as given in Table 7.10.1-3.
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
1s1.0 Gbit/s1.0Gbit/s132 MByte132 MByteUncom­pressed Fede­rated Lear­ning for image reco­gnition
1s80.88 Mbit/s80.88 Mbit/s10 Mbyte10 Mbyte TBDCom­pressed Fede­rated Lear­ning for image/video pro­cessing
1sTBDTBD10 MByte10 MByteData Transfer Distur­bance in Multi-agent multi-device ML Opera­tions
Up

7.10.2  KPI requirement for direct device connection |R19|  p. 106

The 5G system shall support split AI/ML inference between AI/ML endpoints by leveraging direct device connection with performance requirements as given in Table 7.10.2-1.
Max allowed end-to-end latency
(NOTE 1)		 Payload size (Inter­mediate data size)
(NOTE 1)		 Expe­rienced data rate
(NOTE 1)		 Service area dimension Communi­cation service availa­bility
(NOTE 1)		 Relia­bility
(NOTE 1)		 Remarks
10-100 ms≤ 1.5 Mbyte for each frame ≤ 720 Mbps Proximity-based work task offloading for Remote driving, AR displaying/gaming, remote-controlled robotics, video recognition and One-shot object recognition
10 ms≤ 1.6 MByte (8 bits data format)≤ 1.28 Gbps900 m2 (30 m x 30 m)99.999 %99.99 %Local AI/ML model split on factory robots
10 ms≤ 6.4 Mbyte (32 bits data format)≤ 1.5 GbpsLocal AI/ML model split on factory robots
NOTE 1:
The KPIs in the Table apply to UL data transmission in case of indirect network connection. The 5G system shall support AI/ML model/data distribution and sharing by leveraging direct device connection with performance requirements as given in Table 7.10.2-2.
Max allowed end-to-end la­tency
(NOTE 1)		 Expe­rienced data rate
(NOTE 1)		 Payload size
(NOTE 1)		 Communi­cation service availa­bility
(NOTE 1)		 Remark
1s≤ 1.92 Gbit/s≤ 240 MByte99.9 %AI Model Transfer Management through Direct Device Connec­tion
3s≤ 81.33 Mbyte/s≤ 244 MByte-transfer learning for trajec­tory predic­tion
NOTE 1:
The KPIs in the Table apply to data transmission using direct device connection.
NOTE 2:
The AI/ML model data distri­bution is for a specific appli­cation service.
The 5G system shall support AI/ML model/data distribution and sharing by leveraging direct device connection with performance requirements as given in Table 7.10.2-3.
Payload size
(NOTE 1)		 Maximum latency
(NOTE 1)		 Experienced data rate
(NOTE 1)		 Reliability
(NOTE 1)		 Remark
132 MByte2-3 s≤ 528 Mbit/sDirect device connection assisted Federated Learning (Uncompressed model) Asynchronous Federated Learning via direct device connection
≤ 50 MByte1 s≤ 220 Mbit/s99.99%
NOTE 1:
The KPIs in the Table apply to both UL and DL data transmission in case of indirect network connection.
Up

Up   Top   ToC