Skip to content

Cycles

GPU Rendering

AMD

Experimental support for AMD hardware ray-tracing acceleration on Windows, using HIP RT. (557a245)

  • This improves performance on GPUs with ray tracing acceleration - RX 6000 and RX 7000 series, as well as W6000 and W7000 series workstation GPUs.
  • Driver version 22.40.51.06 for Blender 3.6 Beta needs to be installed, along with enabling HIP RT in Preferences > System.
  • Known limitations:
    • No Linux support, as HIP RT is Windows only still.
    • Degenerate triangles may causes crashes or poor performance.
    • Shadows in hair are not rendering accurately.

AMD GPUs now also support light trees. (d5757a0)

Intel

Hardware ray-tracing acceleration for Intel® Arc™ and Data Center GPUs, using Embree 4. (3f8c995109)

Render time per sample on an Intel® Arc™ A770 GPU

Scene With HW RT Without HW RT
barbershop_interior 0.192991 0.223141
bmw27 0.013019 0.015585
classroom 0.112879 0.139789
fishy_cat 0.014602 0.024707
junkshop 0.101691 0.119027
monster 0.057748 0.067487
pabellon 0.045524 0.066214
sponza 0.023778 0.031841
  • During their first use with oneAPI device, Ambient Occlusion and Bevel nodes will trigger a GPU binaries recompilation that will use around 9GB of memory and take several minutes. Improvements on this may come from future GPU drivers.
  • On Windows, when using drivers \< 101.4644, embree on GPU may crash when rendering duration goes beyond 3 minutes.

Apple

Apple Silicon GPUs now support NanoVDB for Metal, reducing memory usage of volumes. (02c2970983)

Performance

  • Light trees now use less memory and are faster to build, through instancing and multi-threading. (bfd1836861, 23c5e06932)
  • Loading large geometries into Cycles is much faster, meaning rendering can start more quickly after geometry changes or switching to rendered view.
    • Loading large meshes 4-6x faster, and mesh attributes are copied up to 10x faster (or even 60x faster for an extreme example with UV maps) (4bcd59d644, 8d0920ec6d).
    • Loading point clouds can be 9x faster (aef0e72e5a).
    • Loading curves can be 10x faster (ae017b3ab7).

Other Improvements

  • Open Shading Language: support for new standard microfacet closures from MaterialX (dielectric_bsdf, conductor_bsdf, generalized_schlick_bsdf).
  • Byte color attributes are now supported for point clouds and curves (5a86c4cc88).
  • Improved Fresnel handling of the Glass BSDF for better energy preservation and accuracy of results at high roughness. (D17149)
  • Bump mapping for diffuse surfaces was improved. (PR#105776)