Skip to content

GPU and WebGPU

The Rust GPU and WebGPU crates contain device kernels and persistent-buffer paths. Support must be described per backend, device, kernel and execution receipt. A common trait or a CPU reference does not qualify every backend.

The Metal qualification tests exercise device identification, 10,000 particles over ten persistent-buffer ticks and a CPU comparison using 2,048 particles. Retained historical execution output records hardware-test passes for its source and toolchain.

The floating-point comparison accepts absolute error up to 5e-5. A test name containing bit_for_bit does not change that tolerance into exact equality. The particle update and bounded integer DES shader do not establish arbitrary model offload or a complete GPU event scheduler.

Claim Required evidence
Actual GPU execution Device/driver identity, backend, enabled features, successful kernel dispatch and source-bound logs.
Numerical agreement Model-specific exact, tolerance or statistical rule; independent reference and edge cases.
End-to-end acceleration Identical model/workload, optimized Rust CPU baseline, repeated timings and explicit compilation/transfer boundaries.
Persistent or zero-copy data Verified ownership/lifetime and measured movement at the particular boundary; no extrapolation to the whole workflow.
Broad backend support Separate installed hardware/browser acceptance for each advertised backend and platform.

Read GPU benchmark evidence, free testing routes and the capability register. General performance leadership and universal cross-device bitwise parity have not been established by the retained kernel tests.