Skip to main content

Methodology explainer

Blender backends are not interchangeable: why HIP and OptiX results stay apart

Blender renders on different GPU backends. UpgradeOrSave compares only within one backend domain, and this page shows the effect.

By UpgradeOrSave EditorialEditorial update Published

Editorial dates describe the text. They do not date the measurements. Measurement acquisition dates are unavailable in this view.

Blender's Cycles renderer can run on several GPU backends: CUDA and OptiX on NVIDIA cards, HIP on AMD cards, oneAPI on Intel cards and Metal on Apple hardware. Each backend is a different software path to the same scene. That makes Blender a useful rendering benchmark and also a trap for anyone who lines up results from different backends as if they were one test.

Why results stay in separate domains

A backend changes more than the vendor badge. Driver stacks, kernel implementations and feature support differ, and those differences show up in render times independently of the hardware's raw capability. UpgradeOrSave therefore treats each backend family as its own reference domain, with its own frozen reference and its own scale anchored at one hundred. We do not normalise between backends, because there is no measured basis for a conversion factor.

What it looks like in practice

The table below is resolved from the gated public evidence snapshot. It shows, for two well-known cards, which backend domains in that snapshot hold a published result. A card with results in different domains has no shared scale with the other, so we cannot state a percentage difference between them for Blender rendering.

Published Blender rendering results by backend domain
HardwareGPU rendering domainHIP rendering domain
Radeon RX 7900 XTXNo published resultPublished result
GeForce RTX 4090Published resultNo published result

States only. A result in one backend domain is never placed on another domain's scale.

This is why some upgrade comparisons on UpgradeOrSave say there is no shared measured workload, even for two popular cards. The data exists; it simply measures different things.

How to compare across vendors anyway

Within a vendor, Blender rankings are directly useful: they order cards on one backend and one scale. Across vendors, use a workload that runs on a common path for both, and check that both products have a ready result in that same domain. If no such workload is published yet, the honest answer is that the cross-vendor rendering comparison is unknown.

What this means for an upgrade

  • Switching GPU vendor for Blender work cannot be judged from rendering indices on different backends.
  • Compare like with like: same backend, same reference domain.

Limitations and uncertainty

  • The presence table reflects the current published state; results can be added or move into review as evidence changes.
  • Some backends have very few published products, which limits what their rankings can tell you.

Evidence and methodology

Source-rights checks and active research holds apply before evidence reaches this page. An unavailable result is omitted rather than estimated.

No commission-tracked links appear in this article. Any future affiliate links require separate activation and a clear disclosure near the link. Read our disclosure.