2019 Rendering 2026: The default GPU layer

wgpu: the shared GPU layer

One GPU API for Vulkan, Metal, D3D and the browser.

wgpu is a Rust implementation of the WebGPU API that runs on Metal, Vulkan, Direct3D 12, OpenGL and, in a browser, the real WebGPU. Application code is the same everywhere. The first step is always the same: create an instance, ask for an adapter (a physical GPU), then open a device and a queue. The output below is from the Mac that built this book.

wgpu-device/src/bin/one.rs ✔ compiled · pollster 0.4, wgpu 30 · rustc 1.95.0

use book_error::Result;

fn main() -> Result<()> {
    let instance = wgpu::Instance::new(wgpu::InstanceDescriptor::new_without_display_handle());

    // Ask for the best GPU the OS and drivers will give us. Having no GPU is a
    // normal condition, so this is an error value and not a panic.
    let adapter = pollster::block_on(instance.request_adapter(&wgpu::RequestAdapterOptions::default()))?;
    let info = adapter.get_info();
    println!("{} via {:?}", info.name, info.backend);

    // A device is the logical connection; the queue submits work to it.
    let (_device, _queue) = pollster::block_on(adapter.request_device(&wgpu::DeviceDescriptor::default()))?;
    Ok(())
}

output

Intel(R) Iris(TM) Plus Graphics via Metal

The adapter reports which backend it used: here an Intel GPU through Metal. Note the constructor: in wgpu 30 an InstanceDescriptor has no Default; you pick new_without_display_handle() or new_with_display_handle(), so that the OpenGL backend can be told whether a display exists. This is the sort of change that shows up in every major version.

wgpu-device/src/bin/two.rs ✔ compiled · pollster 0.4, wgpu 30 · rustc 1.95.0

fn main() {
    // Metal on macOS, Vulkan/GL on Linux, D3D12 on Windows, WebGPU in a browser.
    let mut desc = wgpu::InstanceDescriptor::new_without_display_handle();
    desc.backends = wgpu::Backends::PRIMARY;
    let instance = wgpu::Instance::new(desc);

    // No usable GPU? Ask for a software adapter instead: the route a
    // "software rendering" fallback takes.
    let options = wgpu::RequestAdapterOptions {
        force_fallback_adapter: true,
        ..Default::default()
    };
    match pollster::block_on(instance.request_adapter(&options)) {
        Ok(a) => println!("fallback: {}", a.get_info().name),
        Err(e) => println!("no software adapter here: {e}"),
    }
}

output

no software adapter here: No suitable graphics adapter found; noop not requested, vulkan support not compiled in, metal had no fallback adapters, dx12 support not compiled in, gl not requested, webgpu support not compiled in

force_fallback_adapter asks for a software implementation, the route a "no GPU available" mode would take. On this Mac none exists and wgpu says exactly why. That is the reason several frameworks keep a CPU renderer (Floem's README describes one built on tiny-skia, and Slint has a software renderer) and a reason software paths keep appearing in this ecosystem. wgpu is now the common GPU layer under egui, Slint, Vello and Floem, and versions 29 and 30 shipped in 2026.