The WebGPU API enables web developers to use the underlying system's GPU (Graphics Processing Unit) to carry out high-performance computations and draw complex images that can be rendered in the browser.
WebGL and the GLSL language used to write WebGL shader code are complex, so several WebGL libraries have been created to make WebGL apps easier to write: Popular examples include Three.js, Babylon.js, and PlayCanvas. Developers have used these tools to build immersive web-based 3D games, music videos, training and modeling tools, VR and AR experiences, and more.
However, WebGL has some fundamental issues that needed addressing:
Since WebGL's release, a new generation of native GPU APIs have appeared — the most popular being Apple's Metal, and The Khronos Group's Vulkan — which provide a multitude of new features. There are no more updates planned to OpenGL (and therefore WebGL), so it won't get any of these new features. WebGPU on the other hand will have new features added to it going forwards.
WebGL is based wholly around the use case of drawing graphics and rendering them to a canvas. It does not handle general-purpose GPU (GPGPU) computations very well. GPGPU computations are becoming more and more important for many different use cases, for example those based on machine learning models.
3D graphics apps are becoming increasingly demanding, both in terms of the number of objects to be rendered simultaneously, and usage of new rendering features.
WebGPU addresses these issues, providing an updated general-purpose architecture compatible with modern GPU APIs, which feels more "webby". It supports graphic rendering, but also has first-class support for GPGPU computations. Rendering of individual objects is significantly cheaper on the CPU side, and it supports modern GPU rendering features such as compute-based particles and post-processing filters like color effects, sharpening, and depth-of-field simulation. In addition, it can handle expensive computations such as culling and skinned model transformation directly on the GPU.
There are several layers of abstraction between a device GPU and a web browser running the WebGPU API. It is useful to understand these as you begin to learn WebGPU:
Physical devices have GPUs. Most devices only have one GPU, but some have more than one. Different GPU types are available:
Integrated GPUs, which live on the same board as the CPU and share its memory.
Discrete GPUs, which live on their own board, separate from the CPU.
Software "GPUs", implemented on the CPU.
Note:
The above diagram assumes a device with only one GPU.
A native GPU API, which is part of the OS (e.g., Metal on macOS), is a programming interface allowing native applications to use the capabilities of the GPU. API instructions are sent to the GPU (and responses received) via a driver. It is possible for a system to have multiple native OS APIs and drivers available to communicate with the GPU, although the above diagram assumes a device with only one native API/driver.
A browser's WebGPU implementation handles communicating with the GPU via a native GPU API driver. A WebGPU adapter effectively represents a physical GPU and driver available on the underlying system, in your code.
A logical device is an abstraction via which a single web app can access GPU capabilities in a compartmentalized way. Logical devices are required to provide multiplexing capabilities. A physical device's GPU is used by many applications and processes concurrently, including potentially many web apps. Each web app needs to be able to access WebGPU in isolation for security and logic reasons.
A logical device — represented by a GPUDevice object instance — is the basis from which a web app accesses all WebGPU functionality. Accessing a device is done as follows:
You access an adapter via the GPU.requestAdapter() method. This method accepts an optional settings object allowing you to request for example a high-performance or low-energy adapter. If this is not included, the device will provide access to the default adapter, which is good enough for most purposes.
A device can be requested via GPUAdapter.requestDevice(). This method also accepts an options object (referred to as a descriptor), which can be used to specify the exact features and limits you want the logical device to have. If this is not included, the supplied device will have a reasonable general-purpose spec that is good enough for most purposes.
Putting this together with some feature detection checks, the above process could be achieved as follows:
js
async function init() {
if (!navigator.gpu) {
throw Error("WebGPU not supported.");
}
const adapter = await navigator.gpu.requestAdapter();
if (!adapter) {
throw Error("Couldn't request WebGPU adapter.");
}
const device = await adapter.requestDevice();
/...
}
A pipeline is a logical structure containing programmable stages that are completed to get your program's work done. WebGPU is currently able to handle two types of pipeline:
A render pipeline renders graphics, typically into a <canvas> element, but it could also render graphics offscreen. It has two main stages:
A vertex stage, in which a vertex shader takes positioning data fed into the GPU and uses it to position a series of vertices in 3D space by applying specified effects like rotation, translation, or perspective. The vertices are then assembled into primitives such as triangles (the basic building block of rendered graphics) and rasterized by the GPU to figure out what pixels each one should cover on the drawing canvas.
A fragment stage, in which a fragment shader computes the color for each pixel covered by the primitives produced by the vertex shader. These computations frequently use inputs such as images (in the form of textures) that provide surface details and the position and color of virtual lights.
A compute pipeline is for general computation. A compute pipeline contains a single compute stage in which a compute shader takes general data, processes it in parallel across a specified number of workgroups, then returns the result in one or more buffers. The buffers can contain any kind of data.
The shaders mentioned above are sets of instructions processed by the GPU. WebGPU shaders are written in a low-level Rust-like language called WebGPU Shader Language (WGSL).
There are several different ways in which you could architect a WebGPU app, but the process will likely contain the following steps:
Create shader modules: Write your shader code in WGSL and package it into one or more shader modules.
Get and configure the canvas context: Get the webgpu context of a <canvas> element and configure it to receive information on what graphics to render from your GPU logical device. This step is not necessary if your app has no graphical output, such as one that only uses compute pipelines.
Create resources containing your data: The data that you want processed by your pipelines needs to be stored in GPU buffers or textures to be accessed by your app.
Create pipelines: Define pipeline descriptors that describe the desired pipelines in detail, including the required data structure, bindings, shaders, and resource layouts, then create pipelines from them. Our basic demos only contain a single pipeline, but non-trivial apps will usually contain multiple pipelines for different purposes.
Create a command encoder that can encode a set of commands to be passed to the GPU to execute.
Create a pass encoder object on which compute/render commands are issued.
Run commands to specify which pipelines to use, what buffer(s) to get the required data from, how many drawing operations to run (in the case of render pipelines), etc.
Finalize the command list and encapsulate it in a command buffer.
Submit the command buffer to the GPU via the logical device's command queue.
In the sections below, we will examine a basic render pipeline demo, to allow you to explore what it requires. Later on, we'll also examine a basic compute pipeline example, looking at how it differs from the render pipeline.
We are using the following shader code. The vertex shader stage (@vertex block) accepts a chunk of data containing a position and a color, positions the vertex according to the given position, interpolates the color, then passes the data along to the fragment shader stage. The fragment shader stage (@fragment block) accepts the data from the vertex shader stage and colors the vertex according to the given color.
Note:
In our demos we are storing our shader code inside a template literal, but you can store it anywhere from which it can easily be retrieved as text to be fed into your WebGPU program. For example, another common practice is to store shaders inside a Node.textContent. The correct mime type to use for WGSL is text/wgsl.
To make your shader code available to WebGPU, you have to put it inside a GPUDevice.createShaderModule() call, passing your shader code as a property inside a descriptor object. For example:
In a render pipeline, we need to specify somewhere to render the graphics to. In this case we are getting a reference to an onscreen <canvas> element then calling GPUCanvasContext instance).
From there, we configure the context with a call to GPUDevice that the rendering information will come from, the format the textures will have, and the alpha mode to use when rendering semi-transparent textures.
Note:
The best practice for determining the texture format is to use the GPU.getPreferredCanvasFormat() method; this selects the most efficient format (either bgra8unorm or rgba8unorm) for the user's device.
Next we will provide our WebGPU program with our data, in a form it can use. Our data is initially provided in a Float32Array, which contains 8 data points for each triangle vertex — X, Y, Z, W for position, and R, G, B, A for color.
However, we've got an issue here. We need to get our data into a GPUBuffer. Behind the scenes, this type of buffer is stored in memory very tightly integrated with the GPU's cores to allow for the desired high performance processing. As a side effect, this memory can't be accessed by processes running on the host system, like the browser.
The GPUDevice.createBuffer(). We give it a size equal to the length of the vertices array so it can contain all the data, and VERTEX and COPY_DST usage flags to indicate that the buffer will be used as a vertex buffer and the destination of copy operations.
js
const vertexBuffer = device.createBuffer({
size: vertices.byteLength, / make it big enough to store vertices in
usage: GPUBufferUsage.VERTEX | GPUBufferUsage.COPY_DST,
});
We could handle getting our data into the GPUBuffer using a mapping operation, like we use in the compute pipeline example to read data from the GPU back to JavaScript. However, in this case we are going to use the handy GPUQueue.writeBuffer() convenience method, which takes as its parameters the buffer to write to, the data source to write from, an offset value for each, and the size of data to write (we've specified the whole length of the array). The browser then works out the most efficient way to handle writing the data.
Now we've got our data into a buffer, the next part of the setup is to actually create our pipeline, ready to be used for rendering.
First of all, we create an object that describes the required layout of our vertex data. This perfectly describes what we saw earlier on in our vertices array and vertex shader stage — each vertex has position and color data. Both are formatted in float32x4 format (which maps to the WGSL vec4<f32> type), and the color data starts at an offset of 16 bytes into each vertex. arrayStride specifies the stride, meaning the number of bytes making up each vertex, and stepMode specifies that the data should be fetched per-vertex.
Next, we create a descriptor object that specifies the configuration of our render pipeline stages. For both the shader stages, we specify the GPUShaderModule that the relevant code can be found in (shaderModule), and the name of the function that acts as the entry point for each stage.
In addition, in the case of the vertex shader stage we provide our vertexBuffers object to provide the expected state of our vertex data. And in the case of our fragment shader stage, we provide an array of color target states that indicate the specified rendering format (this matches the format specified in our canvas context config earlier).
We also specify a primitive object, which in this case just states the type of primitive we will be drawing, and a layout of auto. The layout property defines the layout (structure, purpose, and type) of all the GPU resources (buffers, textures, etc.) used during the execution of the pipeline. In more complex apps, this would take the form of a GPUDevice.createPipelineLayout() (you can see an example in our Basic compute pipeline), which allows the GPU to figure out how to run the pipeline most efficiently ahead of time. However, we are specifying the auto value, which will cause the pipeline to generate an implicit bind group layout based on any bindings defined in the shader code.
Now that all the setup is done, we can actually run a rendering pass and draw something onto our <canvas>. To encode any commands to be later issued to the GPU, you need to create a GPUDevice.createCommandEncoder() call.
Next up we start the rendering pass running by creating a GPUCommandEncoder.beginRenderPass() call. This method takes a descriptor object as a parameter, the only mandatory property of which is a colorAttachments array. In this case, we specify:
That the view should be "cleared" to a specified color once loaded and before any drawing takes place. This is what causes the blue background behind the triangle.
That the value of the current rendering pass should be stored for this color attachment.
Now we can invoke methods of the rendering pass encoder to draw our triangle:
GPURenderPassEncoder.setPipeline() is called with our renderPipeline object as a parameter to specify the pipeline to use for the rendering pass.
GPURenderPassEncoder.setVertexBuffer() is called with our vertexBuffer object as a parameter to act as the data source to pass to the pipeline to render. The first parameter is the slot to set the vertex buffer for, and is a reference to the index of the element in the vertexBuffers array which describes this buffer's layout.
GPURenderPassEncoder.draw() sets the drawing in motion. There is data for three vertices inside our vertexBuffer, so we set a vertex count value of 3 to draw them all.
In our basic compute demo, we get the GPU to calculate some values, store them in an output buffer, copy the data across to a staging buffer, then map that staging buffer so that the data can be read out to JavaScript and logged to the console.
The app follows a similar structure to the basic rendering demo. We create a GPUDevice.createShaderModule() call. The difference here is that our shader code only has one shader stage, a @compute stage:
In this example we create two GPUBuffer instances to handle our data, an output buffer to write the GPU calculation results to at high speed, and a stagingBuffer that we'll copy the output contents to, which can be mapped to allow JavaScript to access the values.
output is specified as a storage buffer that will be the source of a copy operation.
stagingBuffer is specified as a buffer that can be mapped for reading by JavaScript, and will be the destination of a copy operation.
When the pipeline is created, we specify a bind group to use for the pipeline. This involves first creating a GPUDevice.createBindGroupLayout()) that defines the structure and purpose of GPU resources such as buffers that will be used in this pipeline. This layout is used as a template for bind groups to adhere to. In this case we give the pipeline access to a single memory buffer, tied to binding slot 0 (this matches the relevant binding number in our shader code — @binding(0)), usable in the compute stage of the pipeline, and with the buffer's purpose defined as storage.
Next we create a GPUDevice.createBindGroup(). We pass this method call a descriptor object that specifies the bind group layout to base this bind group on, and the details of the variable to bind to the slot defined in the layout. In this case, we are declaring binding 0, and specifying that the output buffer we defined earlier should be bound to it.
Running a compute pass is similar in structure to running a rendering pass, with some different commands. For a start, the pass encoder is created using GPUCommandEncoder.beginComputePass().
When issuing the commands, we specify the pipeline to use in the same way as before, using GPUComputePassEncoder.dispatchWorkgroups() to specify the number of GPU workgroups to use to run the calculations.
/ Copy output buffer to staging buffer
commandEncoder.copyBufferToBuffer(
output,
0, / Source offset
stagingBuffer,
0, / Destination offset
BUFFER_SIZE, / Length, in bytes
);
/ End frame by passing array of command buffers to command queue for execution
device.queue.submit([commandEncoder.finish()]);
Once the output data is available in the stagingBuffer, we use the GPUBuffer.getMappedRange(), copy the data into JavaScript, and then log it to the console. We also unmap the stagingBuffer once we are finished with it.
js
/ map staging buffer to read results back to JS
await stagingBuffer.mapAsync(
GPUMapMode.READ,
0, / Offset
BUFFER_SIZE, / Length, in bytes
);
const copyArrayBuffer = stagingBuffer.getMappedRange(0, BUFFER_SIZE);
const data = copyArrayBuffer.slice();
stagingBuffer.unmap();
console.log(new Float32Array(data));
WebGPU calls are validated asynchronously in the GPU process. If errors are found, the problem call is marked as invalid on the GPU side. If another call is made that relies on the return value of an invalidated call, that object will also be marked as invalid, and so on. For this reason, errors in WebGPU are referred to as "contagious".
Once you are done capturing errors, you can end capture by invoking GPUOutOfMemoryError, or GPUValidationError) describing the first error captured in the scope, or null if no errors were captured.
We have attempted to provide useful information to help you understand why errors are occurring in your WebGPU code in "Validation" sections where appropriate, which list criteria to meet to avoid errors. See for example the GPUDevice.createBindGroup() Validation section. Some of this information is complex; rather than repeat the spec, we have decided to just list error criteria that are:
Non-obvious, for example combinations of descriptor properties that produce validation errors. There is no point telling you to make sure you use the correct descriptor object structure. That is both obvious and vague.
Developer-controlled. Some of the error criteria are purely based on internals and not really relevant to web developers.
You can find more information about WebGPU error handling in the explainer — see WebGPU Error Handling best practices provides useful real-world examples and advice.
Note:
The historic way of handling errors in WebGL is to provide a getError() method to return error information. This is problematic in that it returns errors synchronously, which is bad for performance — each call requires a round-trip to the GPU and requires all previously issued operations to be finished. Its state model is also flat, meaning that errors can leak between unrelated code. The creators of WebGPU were determined to improve on this.
Based on a GPUBindGroupLayout, a GPUBindGroup defines a set of resources to be bound together in a group and how those resources are used in shader stages.
Defines the structure and purpose of related GPU resources such as buffers that will be used in a pipeline, and is used as a template when creating GPUBindGroups.
One of the types of errors surfaced by uncapturederror event. Indicates that an operation failed for a system or implementation-specific reason, even when all validation requirements were satisfied.
One of the types of errors surfaced by uncapturederror event. Describes an application error indicating that an operation did not pass the WebGPU API's validation constraints.
