FARPY LABS · SCOPED COMPATIBILITY EVIDENCE

Blender 4.1.1 compatibility testing on FARPY RTX 3090 workers

Controlled evidence for a narrow set of Blender workloads and failure contracts—not a claim of general Blender support.

Current evidence at a glance

  • 4 of 4 registered RTX 3090 GPUs qualified for isolated Blender 4.1.1 research.
  • 4 positive workload classes completed.
  • 6 negative failure-mode behaviors passed.
  • Broader Blender compatibility remains incomplete.
  • Structural visual review passed for all four reviewed fixtures; global visual correctness remains partial.
  • Production eligibility has not been granted.
  • ORBX remains blocked.

Scope: GPU qualification establishes controlled isolated research execution. It does not establish arbitrary-project compatibility, production readiness, scheduler eligibility, or visual quality.

Successful render evidence

Four positive workload classes completed using Blender 4.1.1 and Cycles OptiX. BLEND-012 additionally repeated the canonical still across three RTX 3090 GPUs on two nodes with identical decoded-pixel hashes.

WorkloadFixture and testRuntime and deviceResultProof referenceKnown limitation
BLEND-BASELINE-VALID-STILL Canonical valid-still fixture; one frame and one PNG output. 1 second; Blender 4.1.1; OptiX; CPU disabled. Render exit code 0. Fixture SHA-256 df2ccf024d929d35b126d887bcd2b891b869042aef7a6bdb714475a13df2638c; output SHA-256 0880438d2a825b035b025dcdd87db76a9f2b9ad9ff889044e6b0ae2fa6444788. Canonical evidence: BLEND-BASELINE-VALID-STILL; fixture and output hashes shown here. Structural visual review passed for this fixture; no independent reference image was available.
BLEND-003 Packed-asset fixture; one frame and one PNG output. 7 seconds; Blender 4.1.1; OptiX; CPU disabled. Render exit code 0. Fixture SHA-256 7d5e61678d57cfbc779ecd9b587fdf7e53a2f0dc0d838de1a088abc19f9e02e6; output SHA-256 017e92dc9c1921cbad64b3d3177d30a53f727d3b0f0c3351b448119b9484a39e. Canonical evidence: BLEND-003; fixture and output hashes shown here. Structural visual review passed for this fixture; no independent reference image was available.
BLEND-002 Valid-animation fixture; three frames and three PNG outputs. 44 seconds elapsed; Blender 4.1.1; OptiX; CPU disabled. Render exit code 0. Fixture SHA-256 f7c95c484e5451c9e77ff7a96fe245474fb7d435fce9fe9cccb6d24ba4015bea; expected and actual output counts both 3. Canonical evidence: BLEND-002; fixture hash and output count shown here. Structural visual review passed; no independent reference animation was available.
BLEND-011 Deterministic repeat PASS Three repeated renders produced one decoded-pixel hash. Canonical evidence: BLEND-011 deterministic-repeat proof. Same-node, same-GPU, same-runtime repeat only; cross-node and cross-GPU evidence is recorded separately as BLEND-012.

Expected rejection evidence

These six tests prove controlled rejection and failure classification behavior. They do not count as successful compatible render workloads and do not increase the positive ready-workload count of four.

WorkloadWorkload classFailure classExpected behaviorFalse successSuggested actionProof reference
BLEND-004MISSING_EXTERNAL_ASSETMISSING_EXTERNAL_ASSETRejected as expected; output not accepted.0UPLOAD_OR_PACK_MISSING_ASSETCanonical readiness update: BLEND-004.
BLEND-005MISSING_EXTERNAL_BLEND_LIBRARYMISSING_EXTERNAL_BLEND_LIBRARYRejected as expected; output not accepted.0UPLOAD_OR_PACK_REQUIRED_BLEND_LIBRARYCanonical readiness update: BLEND-005.
BLEND-006NO_CAMERANO_CAMERARejected as expected; output not accepted.0ADD_AND_SELECT_RENDER_CAMERACanonical readiness update: BLEND-006.
BLEND-007NON_CYCLESUNSUPPORTED_RENDER_ENGINERejected as expected; output not accepted.0CONVERT_SCENE_TO_CYCLESCanonical readiness update: BLEND-007.
BLEND-008CORRUPT_OR_UNREADABLE_BLENDCORRUPT_OR_UNREADABLE_BLENDRejected as expected; rendering did not start and no output was accepted.0UPLOAD_VALID_BLEND_FILECanonical readiness update: BLEND-008.
BLEND-009INVALID_FRAME_RANGEINVALID_FRAME_RANGERejected as expected; invalid frame bounds produced no accepted output.0CORRECT_FRAME_RANGECanonical readiness update: BLEND-009.

GPU execution qualification

All four registered NVIDIA GeForce RTX 3090 GPUs passed direct physical-device qualification for controlled, local, isolated Blender 4.1.1 research. Each qualification selected one OptiX GPU, disabled the CPU and the other GPU, proved activity on the selected device, detected no activity on the excluded device, and confirmed idle state before and after.

The original positive and negative compatibility fixtures ran on one qualified GPU. BLEND-012 separately repeated the canonical still on three additional target GPU runs spanning three unique RTX 3090 GPUs across two nodes.

Methods and runtime provenance

  • Blender 4.1.1 on Linux x86_64, installed as an isolated side-by-side research runtime.
  • Exact Blender binary SHA-256: d5cf60563644ed770c5c13952624a0fa8ad076920bbc524275c0044029e36286.
  • Cycles OptiX where applicable, with CPU disabled for GPU qualification and positive GPU render tests.
  • Exactly one selected GPU; the other GPU excluded from each direct qualification.
  • Idle compute-process checks before and after execution.
  • Runtime and fixture hashes captured for identity and repeatability.
  • Standard output, standard error, GPU telemetry, and postflight process-cleanup evidence captured.
  • No production job, queue claim, money action, receipt, artifact delivery, webhook, worker configuration, scheduler configuration, or service restart was created by the positive workload runs.

Visual review and determinism evidence

Eight rendered images from four positive workload classes received a frozen structural visual review. All four reviewed fixtures passed checks for nonblank output, visible geometry, lighting, obvious corruption, and missing-texture fallback.

  • BLEND-BASELINE-VALID-STILL: visible, nonblank, coherent output.
  • BLEND-002: three complete frames with visible animation progression.
  • BLEND-003: packed colored asset visibly present.
  • BLEND-011: three visually identical outputs with one decoded-pixel hash.

The global visual result remains PARTIAL. No independent reference images were available, pixel-perfect reference matching was not performed, and arbitrary Blender visual correctness was not established.

BLEND-012 repeated the canonical still on three unique RTX 3090 GPUs across two nodes. All three target renders completed and matched the canonical decoded-pixel SHA-256:

7478f289798febe4c141c7e6c4d182f69042b517586fbb608deba9184612ea86

This proves cross-GPU and cross-node determinism only for the exact baseline still fixture, Blender 4.1.1 runtime, OptiX configuration, and tested RTX 3090 environment. It does not prove arbitrary Blender determinism, cross-version determinism, cross-runtime determinism, or cross-driver determinism.

Limits

  • Only the documented fixtures and failure modes were tested.
  • Broader Blender compatibility remains incomplete.
  • Structural visual review passed for the four reviewed positive fixtures; independent reference matching and arbitrary-project visual correctness remain unproven.
  • Production eligibility has not been granted.
  • These results do not establish support for arbitrary Blender projects.
  • ORBX testing remains blocked.
  • Negative tests verify rejection behavior, not rendering capability.
  • GPU qualification applies to isolated Blender 4.1.1 research execution.

Proof references

The identifiers above name the frozen internal evidence bundles used for this report. Raw bundles are not linked because they contain private worker and infrastructure data. Public-safe proof material is available through the approved FARPY proof index; this page publishes only scoped findings and safe hashes.

Canonical readiness update: blender-negative-readiness-update-20260729T002440Z.

Canonical visual-review proof: blender-visual-correctness-review-20260729T041420Z.

Canonical cross-GPU and cross-node proof: blend-012-cross-gpu-node-ingest-20260729T042237Z.

Next research steps

  • Expand visual review with frozen independent reference images and larger inspection outputs.
  • Exercise more representative Cycles project features and dependency patterns.
  • Repeat the remaining positive workload classes across additional qualified GPUs to expand execution coverage beyond the canonical still.
  • Close isolation, instrumentation, stop-condition, and production-control gaps before any production decision.
  • Keep ORBX blocked until its runtime, licensing, fixtures, methods, and isolation controls are separately qualified.

Questions and answers

What Blender version was tested?

Blender 4.1.1 on Linux x86_64, using the exact binary hash published in the methods section.

How many GPUs were qualified?

Four of four registered RTX 3090 GPUs qualified for controlled, isolated Blender 4.1.1 research execution.

What successfully rendered?

Exactly four controlled workload classes: a valid still, a packed-asset still, a three-frame animation, and a deterministic-repeat workload.

What failure modes were tested?

Missing external asset, missing external Blender library, no camera, non-Cycles engine, corrupt or unreadable blend input, and invalid frame range.

Do negative tests count as compatible workloads?

No. They verify rejection and classification behavior and do not increase the positive ready-workload count.

Is arbitrary Blender project support proven?

No. Only the documented fixtures and failure modes were tested; broader Blender compatibility remains incomplete.

Was visual correctness reviewed?

Structural visual review passed for all four reviewed fixtures. The global state is PARTIAL because independent reference-image matching, artistic correctness, and arbitrary-project visual correctness were not proven.

Is this production ready?

No. Production eligibility is NOT_GRANTED.

Is ORBX supported?

No. ORBX remains BLOCKED.