Skip to main content
BennyFurman
Participant
September 9, 2026

After Effects 26.3 — GPU acceleration silently falls back to CPU on RTX 50-series (Blackwell / sm_120): shipped CUDA cubins only cover sm_50–sm_90, no PTX

  • September 9, 2026
  • 0 replies
  • 1 view

After Effects 26.3 — GPU acceleration silently falls back to CPU on RTX 50-series (Blackwell / sm_120): shipped CUDA cubins only cover sm_50–sm_90, no PTX

Summary

On an RTX 5070 Ti (GB203, compute capability 12.0 / sm_120), After Effects 26.3 reports Mercury GPU Acceleration (CUDA) as the active renderer, but no CUDA work is ever dispatched. Every GPU-accelerated effect renders on the CPU instead. There is no error message, no warning, and no entry in any log.

The cause appears to be in the shipped kernel binaries: Support Files\PTX\CUDA\*.cubin contains pre-compiled images for sm_50 through sm_90 only, and the folder contains no .ptx images. With no matching cubin and no PTX to JIT-compile from, the CUDA driver cannot produce a loadable image for sm_120, so kernel loading fails and After Effects falls back to CPU rendering.

Practical impact: renders that completed in roughly 10 minutes on the previous (pre-Blackwell) workstation now take hours on newer, substantially faster hardware.

Environment

Component Value
After Effects 26.3 (June 2026 build)
Media Encoder 26.x
Premiere Pro 26.x
OS Windows 11, build 10.0.26200.9445
GPU 1 (discrete) NVIDIA GeForce RTX 5070 Ti — GB203, compute capability 12.0 (sm_120)
NVIDIA driver 32.0.16.1088 (Studio Driver 610.88 WHQL), dated 22-Jul-2026
GPU 0 (integrated) Intel(R) Graphics — driver 32.0.101.8860
System RAM 96 GB
Graphics environment Hybrid (iGPU + dGPU both enumerated)

Adobe's own GPUSniffer, invoked by Media Encoder, records the environment as:

 
gpu-env : hybrid
gpu0 : NVIDIA GeForce RTX 5070 Ti 32.0.16.1088
gpu1 : Intel(R) Graphics 32.0.101.8860

Root cause evidence

Kernel architecture coverage

Each .cubin in Support Files\PTX\CUDA\ is a CUDA fatbinary containing one ELF image per target architecture. Parsing the embedded images (e_machine == 190 / EM_CUDA, architecture read from the low byte of e_flags) gives an identical result for every kernel sampled:

 
sm_50  sm_52  sm_53  sm_60  sm_61  sm_62  sm_70
sm_72 sm_75 sm_80 sm_86 sm_87 sm_89 sm_90

Highest target: sm_90. sm_120 is absent.

16 of the 273 cubins were sampled, chosen across the alphabetical range so as not to bias toward a single effect family. All 16 returned the identical architecture list:

 
AEFX_Levels.cubin              Flare.cubin                  SecondaryPass4.cubin
Accumulate.cubin GaussianBlur.cubin TestPixelUtils.cubin
BleachBypass.cubin IRIDASGainOffset.cubin ToneMapMain.cubin
ColorsQuad.cubin Motion.cubin fxKuwaharaFilter5x5.cubin
Curves.cubin PrColorSecondaryLayer.cubin fxTinting.cubin
FastBoxBlur.cubin

No PTX fallback

Despite the folder being named PTX, it contains no .ptx files. Contents by extension:

Extension Count Purpose
.cubin 273 Pre-compiled CUDA binaries (architecture-specific)
.clz 276 OpenCL kernels (compiled at runtime)
.cso / .csod 276 HLSL / DirectX compute shaders
.ptx 0

PTX would allow the NVIDIA driver to JIT-compile for any architecture, including ones released after the application was built. Without it, forward compatibility is impossible. sm_90 binaries are not compatible with sm_120 — that is a major-version boundary, not a minor one.

Observed behaviour

Verified in Task Manager → Details, with the GPU and GPU engine columns enabled, during an active render.

Renderer = Mercury GPU Acceleration (CUDA):

 
Adobe Media Encoder   PID 12532   CPU 30   GPU 10   GPU 1 - Copy 1
AfterFX.exe PID 38260 CPU 00 GPU 00 (blank)
AfterFX.exe PID 20380 CPU 07 GPU 00 (blank)

Both After Effects processes show no GPU engine at all. Media Encoder shows GPU 1 because NVENC (fixed-function hardware, unaffected by CUDA kernels) is handling the H.264 encode.

Renderer = Mercury GPU Acceleration (OpenCL):

 
Adobe Media Encoder   PID 12532   CPU 00   GPU 04   GPU 0 - Compute
AfterFX.exe PID 20380 CPU 08 GPU 11 GPU 0 - Compute

Under OpenCL a compute engine is engaged — but on GPU 0, the Intel integrated graphics, not the RTX 5070 Ti. This is a separate issue: OpenCL device selection ignores the Windows per-application GPU preference, which is set to the NVIDIA card for every Adobe executable.

Minimal reproduction

  1. New project, 1080p composition, single solid.
  2. Apply Gaussian Blur (a GPU-accelerated effect) with a large radius.
  3. Project Settings → Video Rendering and Effects → Use: Mercury GPU Acceleration (CUDA).
  4. Render ~200 frames.
  5. Task Manager → Details → GPU engine column for AfterFX.exe: blank throughout.

Control: the rest of the pipeline is healthy

Same machine, same GPU, same driver, same week:

Export path Jobs Longest
Premiere Pro timeline → Media Encoder 532 successful 18 minutes
After Effects comp → Media Encoder multiple hours

Premiere Pro is unaffected, which is consistent with it using the HLSL/DirectX compute path plus NVENC rather than the CUDA cubins above. Media Encoder logs confirm hardware encode throughout:

 
- Video: 1080x1920, 25 fps, Progressive, Rec. 709, Hardware Encoding, Nvidia Codec

The Media Encoder error log contains zero GPU, CUDA or NVENC errors across the full retained history.

Already ruled out

  • Clean NVIDIA Studio Driver install; card and driver verified healthy (7 GB VRAM allocated, 50 °C, correct enumeration)
  • Windows Settings → Display → Graphics: GPU preference set to "Specific GPU 1: NVIDIA GeForce RTX 5070 Ti" for AfterFX.exe, aerender.exe, dynamiclinkmanager.exe, GPUSniffer.exe, Adobe Media Encoder.exe, Adobe Premiere Pro.exe
  • Project renderer confirmed set to CUDA in both After Effects and Premiere Pro
  • Multi-Frame Rendering enabled; media cache cleared; preferences reset
  • CPU never saturates during these renders (43–67%), and no disk or network device is saturated — consistent with a serialized CPU fallback rather than a resource limit

Note on AllowUntestedGPUs

The preferences file contains:

 
["GPUManager v2"]
"AllowUntestedGPUs" = 01

This is what allows CUDA to be selected on this card at all. Other users on RTX 50-series report the CUDA option greyed out entirely — the same underlying gap, presented differently. The flag suppresses the compatibility guard without adding kernel coverage, which is why the failure is silent rather than reported.

Requested fix

Ship CUDA kernels that cover Blackwell:

  1. Add sm_120 cubins for the RTX 50-series, and/or
  2. Embed PTX images alongside the cubins so the driver can JIT-compile for architectures released after the build.

Option 2 also prevents this recurring with every future NVIDIA architecture.

Secondary: when no compatible kernel image can be loaded, After Effects should surface an error or fall back visibly, rather than silently CPU-rendering. Users currently have no way to discover this without inspecting the shipped binaries.

Related reports