Compute Architecture for Neural Networks (CANN) is designed for AI tasks on Ascend hardware. It supports AI frameworks such as MindSpore, PyTorch, and TensorFlow. Boosting the performance of AI processors and programming, CANN offers ready-to-use APIs for different uses, allowing you to quickly create AI applications and services with Ascend.
Key Features of CANN 9.1.0
- Added support for Ascend 950DT, providing higher bandwidth in training and decoding inference tasks.
- Added core operators such as SMLA and mHC based on Atlas A2, Atlas A3, and Ascend 950 products. Adapted to open-source models such as DeepSeek V4, Qwen 3.6, Kimi-K2.6, and GLM 5.2.
- Released operator development documentation for beginners, intermediate users, and advanced developers. (Details)
- Based on Ascend 950 products, SHMEM now supports AI Cores directly issuing MTE/UDMA/RDMA data transfers, enabling UB and RoCE communication, and provides matching programming APIs. (Details)
- Based on Ascend 950 products, CATLASS now supports Tile components and provides a full set of MxFP8/MxFP4 quantization templates and sample libraries. (Details)
- The operator precision standard is open-sourced. The precision of floating-point compute operators can now be determined using hybrid tolerance metrics. (Details)
CANN Documentation
View CANN documents by workflow, function, API, and more.