This fork extends the original Yard renderer benchmark into a public results ledger for OpenUSD delivery paths, Hydra render delegates, and VFX Platform stacks.
It records images, process outcomes, timings, memory, logs, and the build stack used. Partial runs and regressions remain useful results when their logs and outputs are preserved.
This is an independent community project. It is not ASWF or OpenUSD certification, and it is not a renderer performance ranking.
Note
Snapshot — 2026-09-07
- Published: 15/15 OpenUSD delivery-path results; five annual Cycles results covering CY2023-CY2027 with patched CY2026/CY2027 follow-ups, including a four-scene patched CY2027 run; the MoonRay CY2025 result with focused texture and shading follow-ups; and five annual Embree results covering CY2023–CY2027 on ASWF prebuilt stacks, plus two CY2027 variants demonstrating the AO-disabled and scene-lit paths. The refreshed CY2027 image also has an updated three-scene Storm result.
- Run model demo:
demo/run1mounts a one-render Storm/Teapot branch into a published image on a GitHub-hosted runner. This demonstrates the branch/image split. - September refresh: released ASWF images repair the tested OSL/OIIO plugin loading and CY2027 Storm/MaterialX failures. Findings from earlier VFX Platform years and focused open questions remain. A complete rerun on the refreshed images is not scheduled.
- Waiting upstream: a tagged Cycles release with the merged Hydra fixes, hdMoonray integration, and OpenUSD Ptex review.
- Not scheduled: modern commercial-delegate coverage.
Published means a pinned snapshot of the run's outputs exists; it does not mean the result was clean. Waiting release has an upstream change but no tagged benchmark stack yet. Waiting upstream is blocked on an external tracker or review. Paused needs a scope, cost, or build review decision. Not scheduled has no current plan.
A benchmark result has two separate inputs:
- A run branch contains the harness, scene selection, Rez package definitions, and the output directories that preserve the result.
- A runnable image contains the OpenUSD and renderer environment used to execute that branch.
Image composition and publishing live in
usd-render-benchmark-stack.
The run branch is mounted at /usr/local/usd-render-benchmark, so its
render_script.sh executes without being copied into the image. Logs, renders,
and the generated summary are written back into the mounted checkout.
The demo/run1
branch is the first published example of this model. Stack workflow
run 34144139279
mounted the branch into the CY2027 runnable image and completed a Storm
Teapot render on a headless GitHub runner.
Each linked label reports successful process exits out of three scenes,
followed by the observed OpenChessSet appearance. Textured, black, and
fallback describe that image, not the other two scenes.
Most links below point to runs published in July and August. The CY2027 prebuilt Storm result was rerun after the September ASWF image refresh.
| Cycle / renderer | Pixar build_usd.py |
ASWF build_usd.sh |
ASWF prebuilt ci-vfxall |
|---|---|---|---|
| CY2023 · GL | 3/3 textured | 3/3 textured | 3/3 fallback |
| CY2024 · GL | 2/3 textured | 2/3 black | 1/3 fallback |
| CY2025 · Storm | 2/3 textured | 2/3 textured | 1/3 fallback |
| CY2026 · Storm | 3/3 textured | 3/3 textured | 0/3 fallback |
| CY2027 · Storm | 3/3 textured | 3/3 fallback | 3/3 textured — refreshed image |
| Delegate | Published results | Current state |
|---|---|---|
| Cycles | Published partial: CY2023 · CY2024 · CY2025 · CY2026 · CY2027 — two of four scenes produce images on each released stack. | Follow-up: patched CY2026 and CY2027 runs add OpenChessSet images after the merged empty-material fix. A newer patched CY2027 four-scene run completes all four renderer processes, including Moana after deferred geometry deletion; its images retain material, texture, and exposure limitations. The changes are under review in Cycles PR #78, with UDIM discovery tracked separately in #77. OptiX-enabled CY2026 and CY2027 images are built, but GPU rendering is not yet established. |
| MoonRay | Published partial: CY2025 — three benchmark images, zero clean exits; Moana not run · tiled-texture and smooth-ALab follow-up | Follow-up: MaterialX BSDF support remains limited (#24); refinement-zero smoothing and missing light-link handling are in hdMoonray #11 and #12. An XPU-capable image exists, but automatic mode used the vector path for these scenes; no XPU benchmark result is claimed. |
| Embree | Published partial: CY2023 · CY2024 · CY2025 · CY2026 · CY2027 — four scenes per stack on ASWF prebuilt (Embree 3.2.2 / 4.3.3); CY2027 renders after the move to OpenUSD 26.08. | Findings: GH #37 CY2027 infeasibility was lifted by OpenUSD 26.08; GH #38 CY2026 texture-read failure; GH #39 black ALab entry — root cause confirmed (AO collapse in enclosed interiors) and closed, with two CY2027 variants published; GH #40 adapter warnings; GH #41 dome-light texture. |
Modern Karma, RenderMan, and Arnold coverage is not scheduled. The original Yard baseline remains the historical mixed-result reference.
| State | Finding | Public record |
|---|---|---|
| Published finding | Material support is limited across delegates. Cycles does not support MaterialX material networks and supports only a subset of UsdPreviewSurface, falling back to its default surface when no supported network is available. MoonRay 2026.29.1 does not support the MaterialX BSDF nodes used by OpenChessSet. |
Cycles #21 · Cycles #25 · MoonRay #24 |
| Published finding | The July–August ASWF CI snapshots did not provide the same working Storm/MaterialX result as OpenUSD built with Pixar's build_usd.py. OpenChessSet rendered textured with the Pixar build, while the corresponding ASWF Conan-based stacks produced fallback or black results with MaterialX errors. |
OpenUSD results · ASWF issues #454 and #455 |
| Published finding | Lighting is not consistent across delegates. The same scene can render with very different exposure and light contribution, so these are stack results rather than look-matched comparisons. | delegate results · Yard baseline |
| Published finding | Exit status does not describe the rendered result on its own. Some runs exit successfully with black or fallback output; others return nonzero after writing a coherent image. | OpenUSD results · delegate results |
| Waiting release | The Cycles empty-material fix is merged; comparable CY2026/CY2027 reruns wait for a tagged release containing it. | Cycles #75 · patched CY2026 · patched CY2027 |
| Waiting upstream | Cycles Hydra diagnostics, AOV reporting, unresolved asset paths, and deferred geometry deletion are under review. UDIM tile discovery remains separate. | Cycles PR #78 · issue #77 · patched CY2027 run |
| Waiting upstream | MoonRay refinement-zero smoothing and missing light-link handling are awaiting upstream integration. | hdMoonray #11 and #12 are open |
| Released improvement | In the refreshed ASWF images, the tested OSL and OIIO discovery plugins load without LD_PRELOAD. The benchmark finding is closed; the ASWF tracking issue remains open. |
Fork issue #3 · aswf-docker #450 |
| Released improvement | Refreshed CY2027 prebuilt Storm renders OpenChessSet with its textured materials on OpenUSD 26.08 / MaterialX 1.39.5, without preload or a MaterialX search-path override. Grey results from earlier VFX Platform years remain tracked separately. | Refreshed CY2027 result · Fork issue #10 · aswf-docker #454 |
| Released improvement | The refreshed CY2027 ASWF build_usd.sh path was reported fixed by its maintainer. The benchmark did not independently rerun that build method. |
Fork issue #2 · aswf-docker #455 |
| Waiting upstream | Storm fails in the Ptex mipmap-loader path on two Moana Island subtrees. | OpenUSD #4168 and #4169 are open; crash-prevention PR #4176 is under review |
Additional run-level findings are tracked in the benchmark issues.
Use the Ideas discussion to suggest delegate/version combinations, available hardware and drivers, licensed-renderer interest, scenes, upstream relevance, and priorities. Failures are valid results, and ASWF reference, Pixar control, and delegate diagnostic runs remain separate.
This phase accepts feedback and coordination only. External code and benchmark
artifact submissions will wait until the MIT license and contribution terms
are restored on main.
The default branch is intentionally a README-only landing page. Each result
link points to a pinned snapshot of the run's outputs whose README describes
the branch and whose render_summary.md, logs/, and renderers/
directories preserve the observed result.
A result README is a starting point, not a guarantee of exact replay. Unknown or historically unrecorded inputs remain limitations rather than being inferred.