The recovered WebGPU surveys covered rendering, texture compression, compute, and mobile performance. They also repeated claims that the new API eliminated driver variation and made performance predictable across devices.
Those claims were too strong. A graphics API gives an application different tools; it does not erase differences in hardware, browser implementation, scene complexity, or thermal conditions.
Verify the renderer you actually ship
Record the library version and the rendering path exercised on each target device. Detect support at runtime and handle unavailable adapters or lost devices.
The Three.js WebGPU renderer guide describes an evolving renderer with important migration differences. Existing shader customizations and post-processing do not become compatible merely because the renderer import changes.
Keep the pipeline guide beside the implementation. Test the fallback as deliberately as the preferred path.
Budget the loaded scene
Count the resources that remain resident after a realistic session: textures, geometry, render targets, shadow maps, and temporary allocations.
A compressed download size is not a GPU-memory measurement. KTX2 is a container; the actual texture format and mip chain determine much of the resident allocation.
Use our texture calculator for an explicit uncompressed RGBA8 estimate. It intentionally does not pretend to measure a particular device or compression target.
Follow the frame through the system
Measure input handling, state updates, scene preparation, rendering, and presentation. A CPU-bound interface may not improve when a shader becomes faster.
Compute can help workloads suited to parallel execution. It also introduces data movement, scheduling, synchronization, and implementation work. Compare the complete operation against its simpler version.
Do not equate every WebGL draw call with a mandatory synchronous CPU-to-GPU round trip. The original prose used that misleading explanation to inflate the contrast.
Inspect the product, not only the benchmark
Check material color, scale, shadows, transparency, labels, and controls. Compare the same camera and lighting before and after renderer or compression changes.
Run a representative sequence: load, customize, resize, change route, return, and repeat. Watch for resources that accumulate or state that is lost.
A mobile check should include the interaction layout and loading experience, not merely an average frame rate.
Add a capability and artifact matrix
The current WebGPU surface includes optional features such as Chrome 152’s subgroup size control. Treat these as separately testable branches. Record whether the test exercised the ordinary compute path or an optimized variant; a success on one cannot silently stand in for the other. Chrome 152.
For each promised device class, keep five results together: initial useful frame, primary interaction, repeated scene replacement, recovery from a failed dependency, and the fallback view. Run a cold visit and a return visit against the same release. A stale browser cache is part of delivery, not an excuse to omit the returning user.
Separate checks that require physical hardware from deterministic checks suitable for CI. A software-rendered screenshot can expose a missing model or wrong label. It cannot establish thermal behavior during ten minutes of touch interaction on a phone.
Bind acceptance to the build output
SLSA 1.2 provenance distinguishes an artifact’s production history from the source revision alone. Apply that distinction to a 3D release: save the built asset-manifest hash beside the application revision and test report.
My suggested release record includes a route, configuration fixture, viewport, backend, asset hash, outcome, and evidence link for every critical flow. Reject the candidate when a required flow is missing evidence, even if all other rows pass. Keep an explicit waiver with an owner and consequence if a product decision accepts a known limitation.
After a rebuild, compare the manifest identity before reusing any screenshot. If it changed, decide which visual and interaction checks the changed assets invalidate. This makes verification selective without making it arbitrary.
Keep the release evidence small and specific
Attach the tested version, devices, key timings, visual captures, and known limits. Describe any quality reduction and the rule that selects it.
The supporting articles in this collection expand each part of the checklist. Together, they replace the archive’s universal speed claims with a more useful standard: the actual application behaves correctly and remains usable on the devices it promises to support.
Keep a good idea close.
Follow Signal Tower
