EXHIBIT 11 / SOFTWARE STACK

What does CANN do between an AI framework and Ascend hardware?

CANN is an Ascend software architecture for AI workloads. It gives frameworks and applications paths to graph compilation, operator libraries, runtime services, communication, and hardware execution; it is neither the model nor the processor.

SHORT ANSWERIt is a software stack that turns framework work into Ascend execution.
6–9 MIN · MANUAL CONTROLS · NO ACCOUNT

ROOM 01 / DEFINE

The stack solves several layers, not one translation step.

  1. 01 / GRAPH Graph Engine

    GE parses, optimizes, compiles, and controls computational graphs, providing a path from framework graphs toward Ascend execution.

  2. 02 / OPS Operator libraries

    Hardware-adapted math, neural-network, computer-vision, and transformer operators supply concrete implementations and API constraints.

  3. 03 / RUNTIME Runtime and application APIs

    Runtime services manage hardware resources and task delivery; application APIs expose model and operator execution capabilities.

  4. 04 / SYSTEM Compiler and communication

    Compilation targets the processor architecture, while communication libraries coordinate collective work across devices.

ROOM 02 / INSPECT

Trace a model port through responsibilities that fail differently.

This sequence combines the official architecture and operator documentation into a reading path. It is not a log from an Ascend machine.

MANUAL EVIDENCE PATHNOT A CAPTURED SYSTEM TRACE

ADAPT

The framework presents a graph

An adapter or graph API maps framework operations and metadata into forms the Ascend graph path understands.

WHAT THIS LETS YOU SAY

A familiar model definition does not remove backend-specific contracts.

Step 1 / 4

Keyboard: focus this instrument and use ← or →. It never advances by itself.

ROOM 03 / TEST

Read a porting status without collapsing the layers.

Select three facts. The verdict identifies the first missing condition; no package is installed and no accelerator is contacted.

LIVE BROWSER CLASSIFIERNO MODEL, HARDWARE, ACCOUNT, OR REMOTE SERVICE

SCENARIOHypothetical framework model moving to an Ascend deployment.

01CANN, driver, firmware, framework plugin

The selected components are outside one verified mapping.

02Required operator cases

The inventory claims coverage for the required cases.

03Numerical and performance record

The model has no attached comparison for this environment.

FIRST FAILURE / ENVIRONMENT

Stop before interpreting model behavior.

An unaligned software stack can invalidate every later compatibility or performance conclusion.

SELECTED FACTS
  • The selected components are outside one verified mapping.
  • The inventory claims coverage for the required cases.
  • The model has no attached comparison for this environment.

CANN releases and mappings change. This exhibit teaches the layers; the official release notes remain authoritative for an actual installation.

ROOM 04 / VERIFY

Use release-scoped documentation, not an ecosystem slogan.

OFFICIAL DOCUMENTATION

A layered architecture

Huawei’s CANN documentation describes framework support, graph development, operator libraries, Runtime and GE APIs, and release mappings.

EDITORIAL TRACE

One porting path

The four-stage sequence is the museum’s synthesis of those responsibilities, not a trace recorded on Ascend hardware.

NOT CLAIMED

Drop-in equivalence

The page does not claim bitwise CUDA equivalence, universal operator coverage, or performance for a model or device.

REVIEWED SOURCES

Open the record before repeating the claim.

CANN Community Edition 9.0.0 documentation 华为昇腾 · key-sections · checked 2026-09-11 The 9.0.0 documentation index plus environment setup, operator, application and graph development, Runtime/GE APIs, and operator-library overviews; every API and hardware model was not reviewed. Operator Registration — PyTorch accelerator documentation PyTorch · key-sections · checked 2026-09-11 Operator set, Dispatch Keys, registration approaches, support-status queries, and troubleshooting sections; every registration API and backend implementation was not reviewed.
NEXT / A COMPLETE AGENT PRODUCTInspect where models, harnesses, channels, and permissions meet →