Build Blender from source on Linux ARM64 with full CUDA GPU support — one command, one hour.
$ blender --version
Blender 5.0.1
$ blender -b --python my_script.py # headless rendering on GB10
There are no official Blender builds for Linux ARM64. Community efforts exist for CPU-only builds, but none support CUDA on the new NVIDIA GB10 (compute capability 12.1) with CUDA 13.
This repo provides a fully automated, one-command build that produces a feature-complete Blender with Cycles GPU rendering on the DGX Spark.
| Component | Version |
|---|---|
| Hardware | NVIDIA DGX Spark / GB10 |
| GPU Compute | SM 12.1 (Blackwell) |
| OS | Ubuntu 24.04 LTS (noble) |
| Arch | aarch64 (ARM64) |
| CUDA | 13.0+ |
| Driver | 580.x+ |
| GCC | 13.3 |
git clone https://github.com/mvalancy/blender-nvidia-gb10.git
cd blender-nvidia-gb10
./setup.sh all # ~1 hour total
blender --versionOr run individual steps:
./setup.sh deps # Install system packages
./setup.sh clone # Clone Blender v5.0.1
./setup.sh patch # Apply ARM64 + CUDA 13 patches
./setup.sh build-deps # Build 60+ third-party libs (~45 min)
./setup.sh build # Build Blender (~10 min)
./setup.sh install # Symlink to /usr/local/binCompleted steps are checkpointed — re-running skips them automatically:
./setup.sh --status # Show which steps are complete vs pending
./setup.sh --force build # Re-run a step (ignoring checkpoint)
./setup.sh --clean # Remove all checkpoints and build artifacts
./setup.sh --verbose all # Show full build output instead of spinner
./setup.sh --help # Full usage infoflowchart TD
A["`**setup.sh all**`"] --> B[deps]
B --> C[clone]
C --> D[patch]
D --> E[build-deps]
E --> F[build]
F --> G[install]
B -->|apt-get| B1[30+ system packages]
B -->|symlink| B2[Ubuntu multiarch fix]
B -->|touch| B3[CUDA 13 header stub]
C -->|git clone| C1["Blender v5.0.1<br>from GitHub"]
C -->|git lfs pull| C2["LFS data files<br>from projects.blender.org"]
D -->|0001-0004| D1["lfdevs ARM64 patches<br>libffi, flex, USD, ROCm"]
D -->|0005-0008| D2["GB10 + CUDA 13 patches<br>OIDN, libglu, Wayland, libdrm"]
E -->|make deps| E1["~60 libraries built"]
F -->|cmake + ninja| F1["Blender binary"]
G -->|symlink| G1["installed to PATH"]
style A fill:#c8e6c9,stroke:#2e7d32,stroke-width:2px,color:#1b5e20
style G1 fill:#c8e6c9,stroke:#2e7d32,stroke-width:2px,color:#1b5e20
style D1 fill:#ffe0b2,stroke:#e65100,color:#bf360c
style D2 fill:#ffe0b2,stroke:#e65100,color:#bf360c
Blender's build system assumes x86 Linux. Eight patches fix the ARM64 + CUDA 13 gaps:
flowchart LR
subgraph lfdevs["Community ARM64 Patches"]
P1["`**0001** libffi<br>GCC 14 missing decl`"]
P2["`**0002** flex<br>GCC 14 segfault`"]
P3["`**0003** USD<br>x86 Valgrind asm`"]
P4["`**0004** ROCm<br>skip on aarch64`"]
end
subgraph gb10["CUDA 13 + Ubuntu 24.04 Patches"]
P5["`**0005** OIDN<br>drop sm_70 Volta`"]
P6["`**0006** libglu<br>libtool mismatch`"]
P7["`**0007** Wayland+Mesa<br>lib64 to lib path`"]
P8["`**0008** FFmpeg<br>link libdrm`"]
end
lfdevs --> gb10
style lfdevs fill:#bbdefb,stroke:#0d47a1,color:#0d47a1
style gb10 fill:#ffcdd2,stroke:#b71c1c,color:#b71c1c
| # | File Modified | Problem | Fix |
|---|---|---|---|
| 0001 | deps/cmake/libffi.cmake |
Missing declaration in libffi with GCC 14 | Add forward declaration |
| 0002 | deps/cmake/flex.cmake |
flex segfaults via reallocarray on GCC 14 |
Patch flex source |
| 0003 | deps/cmake/usd.cmake |
USD includes x86 Valgrind assembly on Linux | Gate behind __x86_64__ |
| 0004 | deps/cmake/ispc.cmake |
ROCm packages are x86-only | Skip on BLENDER_PLATFORM_ARM |
| 0005 | deps/cmake/openimagedenoise.cmake |
CUDA 13 dropped sm_70 (Volta) | Remove sm_70 targets at build time |
| 0006 | deps/cmake/libglu.cmake |
libtool 2.4.6 vs 2.4.7 version mismatch | Run autoreconf -fi before configure |
| 0007 | deps/cmake/wayland.cmake + 4 more |
Meson installs to lib/ not lib64/ on aarch64 |
Conditional harvest paths |
| 0008 | cmake/platform/platform_unix.cmake |
FFmpeg kmsgrab needs libdrm symbols | Add find_library(DRM_LIBRARY drm) |
These are handled automatically by setup.sh deps:
| Issue | Cause | Fix |
|---|---|---|
ISPC can't find bits/wordsize.h |
Ubuntu multiarch puts headers in /usr/include/aarch64-linux-gnu/ |
Symlink bits, gnu, asm to /usr/include/ |
Clang errors on texture_fetch_functions.h |
CUDA 13 removed this header; Clang's wrapper still includes it | Create empty stub file |
| Blender segfaults on startup | Git LFS pointers instead of real data files (131 bytes vs 1MB+) | Pull LFS from projects.blender.org |
graph TB
subgraph build["Build Output"]
BIN["blender binary"]
LIBS["60+ prebuilt libraries"]
end
subgraph runtime["Runtime Stack"]
CYCLES["Cycles Renderer"]
EEVEE["EEVEE"]
PYTHON["Python 3.11"]
end
subgraph hardware["NVIDIA DGX Spark"]
GPU["GB10 GPU<br>SM 12.1, CUDA 13"]
CPU["Grace CPU<br>20 ARM cores"]
RAM["120 GB<br>Unified Memory"]
end
BIN --> CYCLES
BIN --> EEVEE
BIN --> PYTHON
CYCLES -->|CUDA| GPU
CYCLES -->|threads| CPU
EEVEE -->|Vulkan| GPU
PYTHON -->|bpy API| BIN
LIBS --> BIN
style GPU fill:#76b900,stroke:#333,color:#fff,font-weight:bold
style BIN fill:#c8e6c9,stroke:#2e7d32,stroke-width:2px,color:#1b5e20
The primary use case is running Blender headlessly for automated 3D workflows:
# Render a scene on the GPU
blender -b scene.blend -o //output_ -F PNG -f 1 -- --cycles-device CUDA
# Run a Python script
blender -b --factory-startup --python my_pipeline.py
# Detect GPU from Python
blender -b --factory-startup --python-expr "
import bpy
prefs = bpy.context.preferences.addons['cycles'].preferences
prefs.get_devices()
for d in prefs.get_devices_for_type('CUDA'):
print(f'{d.name} (use={d.use})')
"
# Output: NVIDIA GB10 (use=True)After building, run the included test:
blender -b --factory-startup --python test_blender.pyExpected output:
Blender 5.0.1
[CUDA] NVIDIA GB10 (use=True)
Cycles GPU render: PASS
Cycles CPU render: PASS
Python API: PASS
File I/O: PASS
The included benchmark.py creates a procedural Menger sponge fractal (400 glass/metal cubes) with volumetric lighting, then renders at multiple resolutions on the GPU:
blender -b --factory-startup --python benchmark.pyResults on NVIDIA GB10:
| Resolution | Samples | Time |
|---|---|---|
| 480x270 | 16 | 1.6s |
| 1280x720 | 64 | 9.3s |
| 1920x1080 | 128 | 24.9s |
Menger sponge fractal — glass BSDF, metal BSDF, emission, volumetric scatter — rendered on the GB10 via Cycles CUDA.
All renders generated headlessly on the GB10 via blender -b --factory-startup --python <script>. Source scripts in renders/.
 |
 |
| Golden Fibonacci Spiral — 300 metallic + SSS spheres, 384 samples | Glass Sierpinski Tetrahedron — 256 emissive glass tetrahedra, 512 samples |
 |
 |
| Crystal Cave — 42 hexagonal crystals on polished obsidian, 384 samples | Infinite Mirror Corridor — 64-bounce neon reflections, 512 samples |
| Variable | Default | Description |
|---|---|---|
BLENDER_BUILD_DIR |
~/blender-gb10-build |
Override the build/source/checkpoint directory |
NO_COLOR |
(unset) | Disable colored output (also auto-disabled when piped) |
export BLENDER_BUILD_DIR=/path/to/build
./setup.sh all| Resource | Requirement |
|---|---|
| Disk space | ~50 GB (source + deps + build) |
| Build time (deps) | ~45 minutes |
| Build time (Blender) | ~10 minutes |
| RAM | 8 GB minimum, 16+ GB recommended |
| Internet | ~5 GB download (source + deps) |
- lfdevs/blender-linux-arm64 — Community ARM64 CI builds and patches 0001-0004
- CoconutMacaroon/blender-arm64 — DGX Spark build reference
- Blender Developer Forum — Community build guidance
- Blender — The amazing open-source 3D suite
The patches in this repository are provided under the same license as Blender (GPL-3.0). Blender itself is not included — it is cloned from the official repository during the build.
