Architecture

How is the architecture of WebNN?

The WebNN API enables web browsers to leverage native machine learning capabilities provided by the underlying operating system. By implementing a hardware-agnostic abstraction layer, the API allows JavaScript frameworks to access advanced machine learning features without depending on platform-specific implementations. Neural networks in WebNN are represented as computational graphs composed of mathematical operations, which form the foundation for computer vision, natural language processing, and other machine learning applications.

Architecture

What operating systems does WebNN support?

WebNN currently supports Windows, Android, ChromeOS, and macOS. The availability of WebNN operations may vary depending on the device type (CPU, GPU, NPU) and operating system. Progress is ongoing, so your patience is appreciated.

See also: WebNN Browser Compatibility

What WebNN backends are currently supported?

Find WebNN backends support details: LiteRT, DirectML and Core ML

Can I use WebNN API on PC desktop?

The WebNN API runs on common computing devices and operating systems. Performance testing shows effective operation on:

• Current-generation laptops (e.g., devices with Intel Core Ultra mobile processors)
• Modern smartphones (e.g., Google Pixel 3 class devices)

The API is platform-independent and integrates with major ML acceleration frameworks:

• Windows: Windows ML (ONNX Runtime)
• ChromeOS: LiteRT
• Linux: LiteRT
• Android: LiteRT
• Apple platforms: Core ML

The API automatically leverages available hardware acceleration features, including CPU parallel processing, GPU compute capabilities, and dedicated ML accelerators, while maintaining hardware abstraction. Developers can tune performance parameters without managing specific hardware details.

See also: WebNN Browser Compatibility

What is the right level of abstraction for neural network operations?

Neural network operations are mathematical functions used in machine learning frameworks. Modern frameworks support approximately 100 standard operations (convolution, matrix multiplication, reductions, normalizations) with some frameworks offering additional variants.

When designing WebNN operations, we considered decomposing high-level functions into basic mathematical operations. While this would reduce the number of defined operations, it would make networks more verbose and risk losing hardware-level optimizations. Modern platforms provide built-in optimizations for common operations like convolutions and recurrent networks.

WebNN therefore implements both the standard functions and all the smaller operations making up the functions in the spec. This preserves access to platform-specific optimizations while enabling construction of new functions using decomposed operations.

What alternatives have been considered?

WebGPU provides a web API abstraction over graphics hardware that can be used to implement GPU-accelerated neural network operations. Major JavaScript ML frameworks like LiteRT and ONNX Runtime Web currently use WebGPU as a backend. Instead of adopting WebNN, one alternative is to continue using these graphics-based frameworks for ML workloads on the web.

However, this approach has two significant limitations:

1. Limited Hardware Optimization

WebGPU’s graphics abstraction layer cannot fully utilize hardware-specific optimizations and special instructions available at the OS level. While GPU manufacturers continue to innovate with ML-specific hardware features, these performance improvements may not be accessible through generic graphics pipeline interfaces.

2. Conformance Challenges

The diverse hardware ecosystem and multiple driver versions make it difficult to ensure consistent and accurate results across different devices when implementing neural network operations through graphics frameworks. Operating systems traditionally excel at hardware conformance testing and quality assurance - crucial capabilities for ML frameworks, especially in critical applications like healthcare and industrial processes where result accuracy is paramount.

When should web developers use WebNN instead of WebGPU + Wasm?

How do I determine if I am running on Windows ML / OpenVINO Execution Provider on Windows 11?

You can verify which backend you are using by following these steps:

1. Visit https://webnnreport.org/ 
2. Navigate to chrome://histograms/ and search WebNN

WebNN.ORT.WinAppRuntimeInstallState

Find WebNN.ORT.WinAppRuntimeInstallState and check the status code:

  0: Windows Version Too Old    5: Paused 
  1: Activation Failure         6: Paused Low Battery
  2: Completed                  7: Paused WiFi Recommended
  3: Error                      8: Paused WiFi Required
  4: Canceled                   9: Runtime Already Present

0  ... 
9  -O          (1 = 100.0%) {0.0%}  // Already installed
10 ... 

WebNN.ORT.OpenVINOExecutionProvider.Status

Find WebNN.ORT.OpenVINOExecutionProvider.Status and check the status code:

  0: Unknown                    3: Ensure ready failed
  1: EP version too low         4: Ready for use
  2: Not installed

  0  ... 
  2  -O          (1 = 50.0%) {0.0%}
  3  O           (0 = 0.0%) {50.0%}
  4  -O          (1 = 50.0%) {50.0%} // Installed
  5  O           (0 = 0.0%) {100.0%}

If the Windows ML Runtime and OpenVINO Execution Provider failed to install properly, you can install them manually instead.

Windows App SDK and Execution Providers Manual Installation

On Window 11, version 24H2 or newer, the Windows App SDK and Execution Providers will be installed automatically in the background when you install and launch the browser. The installation time may vary depending on your network speed. Follow these steps if the Windows App SDK and Execution Providers were not installed successfully.

C:\Users\webnn\Downloads>ExecutionProviderCatalog.exe

Creating dynamic dependency on Microsoft.WindowsAppRuntime.1.8_8wekyb3d8bbwe
Successfully created package dependency with ID: T:XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX
Successfully added package dependency. Package: Microsoft.WindowsAppRuntime.1.8_8000.625.330.0_x64__8wekyb3d8bbwe
Package dependency will be automatically cleaned up when process exits
Got default ExecutionProviderCatalog
Found 1 providers
Provider 1:
    Name: OpenVINOExecutionProvider
    Calling EnsureReadyAsync...
    EnsureReadyAsync started successfully
    Waiting for EnsureReadyAsync to complete...
    [SUCCESS] Provider ready successfully
    Library Path: C:\Program Files\WindowsApps\MicrosoftCorporationII.WinML.Intel.OpenVINO.EP.1.8_1.8.15.0_x64__8wekyb3d8bbwe\ExecutionProvider\onnxruntime_providers_openvino_plugin.dll