What's New in WebGPU (Chrome 134)

François Beaufort

Published: February 26, 2025

Improve machine-learning workloads with subgroups

After a year of development and trials, the subgroups WebGPU feature enabling SIMD-level parallelism is now available. It allows threads in a workgroup to communicate and execute collective math operations, such as calculating a sum of numbers, and offers an efficient method for cross-thread data sharing. See the original proposal and chromestatus entry.

For reference, Google Meet saw 2.3-2.9 times speed increases when benchmarking subgroups against packed integer dot products for matrix-vector multiply shaders on some devices during the origin trial.

When the "subgroups" feature is available in a GPUAdapter, request a GPUDevice with this feature to get subgroups support in WGSL. It's helpful to check subgroupMinSize and subgroupMaxSize adapter info values—for example, if you have a hardcoded algorithm that requires a subgroup of a certain size.

You also need to explicitly enable this extension in your WGSL code with enable subgroups; to get access to the following built-in values in both compute and fragment shaders stages:

• subgroup_invocation_id: A built-in value for the index of the thread within the subgroup.

• subgroup_size: A built-in value for subgroup size access.

The numerous subgroup built-in functions (for example, subgroupAdd(), subgroupBallot(), subgroupBroadcast(), subgroupShuffle()) enable efficient communication and computation between invocations within a subgroup. These subgroup operations are classified as single-instruction multiple-thread (SIMT) operations. Additionally, the quad built-in functions, which operate on a quad of invocations facilitate data communication within the quad.

You can use f16 values with subgroups when you request a GPUDevice with both "shader-f16" and "subgroups" features.

The following sample is a good starting point for exploring subgroups: it shows a shader that uses the subgroupExclusiveMul() built-in function to compute factorials without reading or writing memory to communicate intermediate results.

Remove float filterable texture types support as blendable

Now that the 32-bit float textures blending is available with the "float32-blendable" feature, the incorrect support for float filterable texture types as blendable is removed. See issue 364987733.

Dawn updates

Dawn now requires macOS 11 and iOS 14 and only supports Metal 2.3+. See issue 381117827.

The new GetWGSLLanguageFeatures() method of the wgpu::Instance now replaces EnumerateWGSLLanguageFeatures(). See issue 368672124.

The following binding types have an Undefined value and their default values in binding layout have been changed. See issue 377820810.

• wgpu::BufferBindingType::Undefined is now Uniform
• wgpu::SamplerBindingType::Undefined is now Filtering
• wgpu::TextureSampleType::Undefined is now Float
• wgpu::StorageTextureAccess::Undefinedis now WriteOnly

This covers only some of the key highlights. Check out the exhaustive list of commits.

What's New in WebGPU

A list of everything that has been covered in the What's New in WebGPU series.

Chrome 134

• Improve machine-learning workloads with subgroups
• Remove float filterable texture types support as blendable
• Dawn updates

Chrome 133

• Additional unorm8x4-bgra and 1-component vertex formats
• Allow unknown limits to be requested with undefined value
• WGSL alignment rules changes
• WGSL performance gains with discard
• Use VideoFrame displaySize for external textures
• Handle images with non-default orientations using copyExternalImageToTexture
• Improving developer experience
• Enable compatibility mode with featureLevel
• Experimental subgroup features cleanup
• Deprecate maxInterStageShaderComponents limit
• Dawn updates

Chrome 132

• Texture view usage
• 32-bit float textures blending
• GPUDevice adapterInfo attribute
• Configuring canvas context with invalid format throw JavaScript error
• Filtering sampler restrictions on textures
• Extended subgroups experimentation
• Improving developer experience
• Experimental support for 16-bit normalized texture formats
• Dawn updates

Chrome 131

• Clip distances in WGSL
• GPUCanvasContext getConfiguration()
• Point and line primitives must not have depth bias
• Inclusive scan built-in functions for subgroups
• Experimental support for multi-draw indirect
• Shader module compilation option strict math
• Remove GPUAdapter requestAdapterInfo()
• Dawn updates

Chrome 130

• Dual source blending
• Shader compilation time improvements on Metal
• Deprecation of GPUAdapter requestAdapterInfo()
• Dawn updates

Chrome 129

• HDR support with canvas tone mapping mode
• Expanded subgroups support
• Dawn updates

Chrome 128

• Experimenting with subgroups
• Deprecate setting depth bias for lines and points
• Hide uncaptured error DevTools warning if preventDefault
• WGSL interpolate sampling first and either
• Dawn updates

Chrome 127

• Experimental support for OpenGL ES on Android
• GPUAdapter info attribute
• WebAssembly interop improvements
• Improved command encoder errors
• Dawn updates

Chrome 126

• Increase maxTextureArrayLayers limit
• Buffer upload optimization for Vulkan backend
• Shader compilation time improvements
• Submitted command buffers must be unique
• Dawn updates

Chrome 125

• Subgroups (feature in development)
• Render to slice of 3D texture
• Dawn updates

Chrome 124

• Read-only and read-write storage textures
• Service workers and shared workers support
• New adapter information attributes
• Bug fixes
• Dawn updates

Chrome 123

• DP4a built-in functions support in WGSL
• Unrestricted pointer parameters in WGSL
• Syntax sugar for dereferencing composites in WGSL
• Separate read-only state for stencil and depth aspects
• Dawn updates

Chrome 122

• Expand reach with compatibility mode (feature in development)
• Increase maxVertexAttributes limit
• Dawn updates

Chrome 121

• Support WebGPU on Android
• Use DXC instead of FXC for shader compilation on Windows
• Timestamp queries in compute and render passes
• Default entry points to shader modules
• Support display-p3 as GPUExternalTexture color space
• Memory heaps info
• Dawn updates

Chrome 120

• Support for 16-bit floating-point values in WGSL
• Push the limits
• Changes to depth-stencil state
• Adapter information updates
• Timestamp queries quantization
• Spring-cleaning features

Chrome 119

• Filterable 32-bit float textures
• unorm10-10-10-2 vertex format
• rgb10a2uint texture format
• Dawn updates

Chrome 118

• HTMLImageElement and ImageData support in copyExternalImageToTexture()
• Experimental support for read-write and read-only storage texture
• Dawn updates

Chrome 117

• Unset vertex buffer
• Unset bind group
• Silence errors from async pipeline creation when device is lost
• SPIR-V shader module creation updates
• Improving developer experience
• Caching pipelines with automatically generated layout
• Dawn updates

Chrome 116

• WebCodecs integration
• Lost device returned by GPUAdapter requestDevice()
• Keep video playback smooth if importExternalTexture() is called
• Spec conformance
• Improving developer experience
• Dawn updates

Chrome 115

• Supported WGSL language extensions
• Experimental support for Direct3D 11
• Get discrete GPU by default on AC power
• Improving developer experience
• Dawn updates

Chrome 114

• Optimize JavaScript
• getCurrentTexture() on unconfigured canvas throws InvalidStateError
• WGSL updates
• Dawn updates

Chrome 113

• Use WebCodecs VideoFrame source in importExternalTexture()