cann-robotupdate v9.1.0-beta2 release note

CHANGELOG

v9.1.0-beta.2

Release Date: 2026/05/31

CANN Version	Version Source Tag
CANN 9.1.0-beta.2	v9.1.0-beta.2

🚀 Key Features

The AscendC framework basic API supports NPU Check (PR#1557 PR#1467), enhancing operator runtime verification capability.
SIMD VF supports printf and reg dump printing (PR#1605), providing debug printing and register data dump capability.
A5 supports DumpTensor for L1 Tensor data (PR#2175), extending L1 layer data debugging support.
The compilation project CMakeModule supports CMAKE compilation options (PR#2055); adds the optype_collector tool, supporting optype duplicate name checking (PR#285).
The basic API supports functional behavior in ctrl (saturation overflow management) (PR#2077).
SIMT programming adds ld/st interfaces (PR#2058) and AddrSpace class interfaces (PR#1597), enriching SIMT memory access programming capability.

🎯 Sample Updates

Best practice sample development: matmul+gelu fusion, datacopy optimization, bank conflict optimization, group_matmul quantized group matrix multiplication, and simt&simd high-performance programming (PR#1814 PR#2137 PR#2141 PR#2166 PR#2363).
Ascend 950 new feature supplements and compatibility sample rectification: loopmode data movement, interleave vector computation, datacopy_gm2l1, loadmx (Load2DMX), mmad_mx, data_copy_pad, and so on (PR#2336 PR#1899 PR#2124).
RegBase adds basic samples: basic arithmetic, data type conversion, reduction, comparison, indexing, and other samples (PR#1459 PR#1575 PR#2024).
Added SIMD VF print samples and dump samples (PR#2558).
SIMT adds DCache access optimization samples (PR#2453), transpose-based memory coalescing and bank conflict samples (PR#1753), and best practice samples for improving data movement efficiency through type alignment (PR#2297).
SIMT adds functional feature samples: PyTorch custom operator registration (PR#2769), compilation-related samples (dynamic, static, separate compilation, and so on) (PR#2356), profiling samples (PR#1989), memory barrier feature samples (PR#1923), Warp class feature samples (PR#2876), simulator samples (PR#2692), and kernel log samples (PR#2131).
The SIMT getting-started sample is modified to gather (PR#2405).
Added Tensor API getting-started and best practice samples: Matmul getting started, data movement in/out, data movement out with on-path quantization, and MX FP4 best practices (PR#2553).

📖 Documentation

Added matrix computation overview and computation fractal introduction documentation (PR#2533).
Optimized vector computation API documentation, supplementing instruction constraints, and so on (PR#2676).
Set up a VitePress documentation site, providing AscendC documentation preview functionality (PR#2547).
Added SIMD and SIMT hybrid programming performance optimization overview (PR#2736).

v9.1.0-beta.1

Release Date: 2026/04/30

CANN Version	Version Source Tag
CANN 9.1.0-beta.1	v9.1.0-beta.1

🚀 Key Features

Aclrtc supports global functions as template parameters (PR#1276), and fusion compilation supports msobjdump parsing (PR#235), improving compilation project flexibility and debugging capability.
Tensor API branch code is merged into master (PR#1770), officially providing Tensor programming support.
C-API adds support for moving int4x2_t type data between UB and registers (PR#1404 PR#1504), extending low-bit data movement capability.
SIMT API adds DCCI interfaces (PR#1510), floating-point comparison interfaces (PR#1693), and floating-point operation instructions (PR#1774), enriching SIMT programming capability.

🎯 Samples

Existing sample rectification: rectification of basic, high-level cube, vector, and data movement API samples (PR#1200 PR#1429 PR#1454 PR#1311 PR#1563).
Added static Tensor best practices: add high-performance sample (PR#1262).
Added low-bit cube best practices: matmul high-level and basic API samples, MXFP4 basic API samples (PR#1535 PR#1648 PR#1788).
Added custom operator project multi-vendors compilation sample (PR#1466).
Added atomic feature samples (PR#1688).

📖 Documentation

Added C-API manual content: supplemented UB and register data movement interface documentation (PR#1404 PR#1504).
Optimized getting-started tutorials and SIMT programming model (PR#1487).
Added CANN package installation and compatibility instructions (PR#1451 PR#1514).
Added AscendC getting-started tutorials, programming guides, compatibility migration guides, and operator practice reference documentation (PR#1438).

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.