Meeting: TSGS4_135-bis-e | Agenda Item: 7.8
Proposals for ULBC Codec Complexity Measurement
Cadence Design Systems Inc.
discussion
Discussion
This contribution presents feedback on complexity measurement methodology for the ULBC from an implementor's perspective. Current complexity metrics focused primarily on MAC operations and NN Model-size. These are insufficient to capture the full computational burden of AI-based codecs on real embedded hardware. A more comprehensive framework is proposed that accounts for four distinct operator categories - Data Reorganization, NN MAC-Core-Operations, Vector ALU operations, and Activation Functions - along with separate memory complexity metrics covering maximum required static weight storage, peak and average dynamic scratch memory, along with total NN Model size in bytes. The contribution further highlights the importance of considering diverse hardware platforms ranging from low-power DSPs to high-performance DSPs with NN extensions and dedicated NPUs, each offering significantly different MAC throughput and memory bandwidth characteristics. With this contribution, we propose that the 3GPP SA4 working group align on a simple yet robust complexity measurement approach that enables fair comparison between ULBC and legacy codecs such as EVS and IVAS and enables close resemblance to actual on-device complexity through use of open-source AI frameworks
| TDoc | S4-260527 |
| Title | Proposals for ULBC Codec Complexity Measurement |
| Source | Cadence Design Systems Inc. |
| Agenda item | 7.8 |
| Agenda item description | FS_ULBC (Study on Ultra Low Bitrate Speech Codec) |
| Doc type | discussion |
| For action | Discussion |
| Abstract | This contribution presents feedback on complexity measurement methodology for the ULBC from an implementor's perspective. Current complexity metrics focused primarily on MAC operations and NN Model-size. These are insufficient to capture the full computational burden of AI-based codecs on real embedded hardware. A more comprehensive framework is proposed that accounts for four distinct operator categories - Data Reorganization, NN MAC-Core-Operations, Vector ALU operations, and Activation Functions - along with separate memory complexity metrics covering maximum required static weight storage, peak and average dynamic scratch memory, along with total NN Model size in bytes. The contribution further highlights the importance of considering diverse hardware platforms ranging from low-power DSPs to high-performance DSPs with NN extensions and dedicated NPUs, each offering significantly different MAC throughput and memory bandwidth characteristics. With this contribution, we propose that the 3GPP SA4 working group align on a simple yet robust complexity measurement approach that enables fair comparison between ULBC and legacy codecs such as EVS and IVAS and enables close resemblance to actual on-device complexity through use of open-source AI frameworks |
| download_url | https://www.3gpp.org/ftp/tsg_sa/WG4_CODEC/TSGS4_135-bis-e/Docs/S4-260527.zip |
| For | Discussion |
| Type | discussion |
| Contact | Satish Patil |
| Contact ID | 105880 |
| TDoc Status | reserved |
| Reservation date | 4/6/2026 9:35:45 AM |
| Agenda item sort order | 18 |