CHANGELOG
v9.0.0-beta.2
Release Date: 2026/03/31
| CANN Version | Version Source Tag |
|---|---|
| CANN 9.0.0-beta.2 | v9.0.0-beta.2 |
🚀 Key Features
- Ascend 950PR supports SIMD programming mode, providing 200+ API interfaces with cross-generation compatibility, enabling smooth operator migration from Atlas A2 series and Atlas A3 series products.
- Ascend 950PR adds Reg-based programming, providing 90+ Reg programming interfaces including Reg data movement, basic arithmetic, reduction computation, and synchronization control.
- Atlas A2 series, Atlas A3 series, and Ascend 950PR support language extension layer pure C interfaces, providing array-style memory allocation and pointer-based computation interfaces for native pure C programming experience.
- Ascend 950PR supports SIMD and SIMT hybrid programming, providing approximately 700 SIMT API interfaces, including warp, atomic, basic mathematical computation, type conversion, and other fundamental interfaces.
- Ascend 950PR supports communication high-level API CCU communication interfaces, providing Allreduce, Allgather, Reducescatter, AlltoAll and other mainstream communication primitives based on CCU; Matmul high-level API adds support for MXFP4/8 low-bit data type matrix operations, achieving half memory usage and doubled computing throughput.
- Ascend 950PR adds and compatibly supports approximately 260 samples, including SIMT samples, SIMD samples (framework, basic API, high-level API, best practices, and so on), and sample directory structure adjustments to improve readability.
- Fusion compilation and <<<>>> invocation support CPU mode and SIM simulation mode.
📖 Documentation
- Added 90+ Reg programming interface API documentation. Reg vector computation API is developed for RegBase architecture, allowing users to directly operate registers involved in Vector computation on the chip for greater flexibility and better performance.
- Added SIMT quick start, programming model, and operator implementation introductions.
- Added SIMD and SIMT hybrid programming model, operator implementation, and performance optimization introductions.
- Added SIMT API documentation section.
- Added compatibility migration guide (220x architecture version migration to 351x architecture version).
- In Ascend community, Ascend C operator development added visualization zone, presenting Cube and Vector operator execution processes through videos.
v9.0.0-beta.1
Release Date: 2026/02/25
| CANN Version | Version Source Tag |
|---|---|
| CANN 9.0.0-beta.1 | v9.0.0-beta.1 |
🚀 Key Features
- Completed migration of Gitee sample repository AscendC samples and AscendC high-level API repository samples to asc-devkit repository, and unified <<<>>> invocation method;
- Fusion compilation and <<<>>> invocation support CPU mode and SIM simulation mode.
- Ascend C project adds support for Kirin X90 and Kirin 9030 processors. Harmony developers can refer to cann-recipes-harmony-infer for Ascend C practice cases.
v8.5.0
Release Date: 2026/01/23
| CANN Version | Version Source Tag |
|---|---|
| CANN 8.5.0 | v8.5.0 |
🚀 Key Features
Based on Atlas A3 training series products/Atlas A3 inference series products, Atlas A2 training series products/Atlas A2 inference series products fully open source, including the following new features:
- Implemented repository separation and packaging, supporting independent installation and deployment of separate packages, including core repository asc-devkit for Ascend C operator development, debugging tool repository asc-tools, Vector operator template library repository atvc and atvoss, Python frontend repository pyasc.
- Programming API capability expansion
- Added language extension layer C API, providing industry-similar programming experience.
- Basic API added LocalMemAllocator memory allocation interface.
- Full support for heterogeneous compilation and <<<>>> direct invocation, enabled through file suffix ".asc" or compilation option "-x asc".
- Standardized operator compilation CMake interfaces, providing Cmake module interfaces supporting different compilation scenarios.
- Supported CPU&NPU twin debugging capability consistency, one set of code supporting both CPU and NPU debugging.
- Comprehensive optimization of programming guide.
- Added operator samples.