Browse categories
Explore
Fiverr Pro
English
$
USD
I will compile and optimize mediapipe for your arm device with GPU acceleration
Richter
About this gig
MediaPipe doesn't ship ARM64 wheels. I build them with GPU acceleration.
I compile from Bazel source, patched for ARM Mali GPU with EGL/GBM headless support. You get a pip-installable .whl with GPU delegate working no X11, no display server, no Docker GPU headaches.
What you get:
Custom .whl for your ARM board + Python + MediaPipe version
GPU delegate via EGL GBM (truly headless)
Install script + verification test
Benchmark report (CPU vs GPU, latency + throughput)
Verified platforms:
RK3576 (Mali-G52) primary dev board
RK3588 (Mali-G610)
Raspberry Pi 5 (VideoCore VII)
Any ARM64 Linux with Mali/VideoCore GPU + DDK
Benchmark: https://asciinema.org/a/Mv4LEGvaroBSs6oJ
Why this matters:
Stock: CPU-only, 100+ms/frame on ARM
My build: GPU-accelerated, 44ms/frame (2.3x faster)
Headless: Docker, CI/CD, server rack
No NPU SDK needed standard GPU drivers only
What I need:
Board model + OS (Ubuntu, Debian, Yocto)
Python version (3.10/3.11/3.12)
Modules: Pose, Face, Hand, Holistic, or all
Contact me before ordering if your setup is unusual I'll confirm compatibility.
Get to know Richter
Richter
- FromChina
- Member sinceOct 2024
- Last delivery1 year
Languages
English, Chinese, German
I build computer vision systems that ship — on NVIDIA CUDA servers and ARM edge.
Not demos. Production.
6 projects deployed in 12 months: YOLO detection + tracking on CUDA and NPU (17x speedup), multi-camera RTSP pipelines with FFmpeg hardware decoding, MediaPipe GPU compiled from source for ARM Mali (2.3x faster, headless), PyTorch custom model training, and rPPG contactless vital signs from video.
Stack: Python, C++, PyTorch, OpenCV, CUDA, ONNX, YOLO, Docker. GPUs: RTX 4060 Ti, Hailo-8L NPU, Mali-G52.
3600+ lines in a real school. 20K+ lines in a shipping edge AI product.
My Portfolio
FAQ
Q: Why can't I just pip install mediapipe on my ARM board?
A: Google only publishes x86_64 wheels. ARM64/aarch64 has no official wheel. You must compile from source using Bazel, which requires ~30GB build space and 1-2 hours. I've already solved the hard parts (EGL/GBM patching, Bazel config for ARM, GPU driver linking).
Q: What's the difference between CPU and GPU build?
A: CPU build uses xnnpack for inference — ~100ms per frame on RK3576. GPU build uses Mali GPU via EGL/OpenGLES — ~44ms per frame. Same accuracy, same model, 2.3x faster. GPU build also frees the CPU for other tasks (video decoding, API serving).
Q: Do you provide the source patches?
A: Premium package includes all Bazel BUILD files, CMake patches, and EGL/GBM modifications as a patch set you can reapply to future MediaPipe versions. Basic and Standard include only the compiled wheel.
Q: Will it work in Docker?
A: Yes. The GPU build links against /dev/dri/renderD128 (DRM render node), which Docker can expose via --device. I provide a tested Dockerfile in Standard and Premium packages.
Q: How long does the build take on my hardware?
A: Compilation happens on MY hardware (I have the toolchain ready). You receive the finished .whl file. Installation on your device takes ~30 seconds via pip install.
