FFmpeg Ampere Neoverse-N1 Optimized Image

This image provides a build of FFmpeg optimized specifically for the ARM Neoverse-N1 architecture, following the Ampere Computing tuning guidelines.

Docker Image: somewatson/ffmpeg-ampere-n1

Source: Ampere FFmpeg Tuning Guide

Build the Image

Build the image from the local Dockerfile:

docker build -t ffmpeg-ampere-n1 .

Note on High Core Counts: This Dockerfile uses $(nproc) to maximize parallelism. On machines with a very high number of cores (e.g., 128+), ensure you have sufficient RAM (roughly 2GB per core) to avoid Out-Of-Memory (OOM) errors during compilation. If the build fails due to memory, you may need to limit the parallelism by replacing $(nproc) with a fixed number in the Dockerfile.

Usage

The image is configured with ffmpeg as the entrypoint.

Check Version

Verify the installation and optimization:

docker run --rm ffmpeg-ampere-n1 -version

Basic Transcoding

To transcode a video file, mount your local media directory to the container. We recommend using --shm-size=2g and --privileged for maximum performance on Ampere N1:

docker run --rm --shm-size=2g --privileged -v $(pwd):/media ffmpeg-ampere-n1 \
  -i /media/input.mp4 \
  -c:v libx264 \
  -preset medium \
  -crf 23 \
  -c:a aac \
  /media/output.mp4

H.265 (HEVC) Encoding

Utilizing the optimized libx265:

docker run --rm --shm-size=2g --privileged -v $(pwd):/media ffmpeg-ampere-n1 \
  -i /media/input.mp4 \
  -c:v libx265 \
  -crf 28 \
  /media/output_hevc.mp4

AV1 Encoding

Utilizing libsvtav1 for high-performance, scalable encoding.

Note: This image uses SVT-AV1 instead of the reference libaom-av1 implementation. SVT-AV1 is specifically designed for massive multi-core parallelism, making it significantly faster and more efficient on high-core-count systems (e.g., 128+ cores).

docker run --rm --shm-size=2g --privileged -v $(pwd):/media ffmpeg-ampere-n1 \
  -i /media/input.mp4 \
  -c:v libsvtav1 \
  -crf 30 \
  -preset 6 \
  /media/output_av1.mp4

Large-Scale Encoding (Chunked Parallelism)

For maximum throughput on high-core-count systems, use chunked encoding. This process splits the input into segments, encodes them in parallel across multiple FFmpeg instances, and concatenates the results.

Use the provided helper script:

chmod +x chunked_encode.sh
./chunked_encode.sh <input_file> <output_file> <codec> <crf> <preset> <chunks> <image_tag>

# Example: Split into 10 chunks using libx265
./chunked_encode.sh input.mp4 output.mp4 libx265 28 slow 10 ffmpeg-ampere-n1

Benchmarks

Performance comparison between a generic FFmpeg image and the optimized ffmpeg-ampere-n1 image on Ampere Neoverse-N1 architecture.

Quality Metric (PSNR): PSNR (Peak Signal-to-Noise Ratio) measures reconstruction quality.

• > 40 dB: Excellent (imperceptible difference from source)
• 30-40 dB: Good to Very Good
• < 30 dB: Noticeable quality loss

Test Configuration:

• Source: BigBuckBunny_512kb.mp4
• Settings: CRF 23, Preset Slow (or Preset 8 for SVT-AV1)
• Runtime: --ipc=host, --privileged

Standard Mode

Image	Codec	Time (s)	Size (KB)	PSNR (dB)	FPS
Generic	libx264	51.84	26,092	41.36	286.00
Optimized	libx264	50.93	26,092	41.36	286.00
Generic	libx265	312.62	17,268	39.44	46.00
Optimized	libx265	253.01	17,420	39.36	57.00
Generic	libsvtav1	36.62	29,188	41.94	412.00
Optimized	libsvtav1	28.82	29,232	41.95	515.00

Standard Mode Performance Improvement:

• libx264: 1.00% faster (Speedup: 0.91s)
• libx265: 19.00% faster (Speedup: 59.61s)
• libsvtav1: 21.00% faster (Speedup: 7.80s)

Chunked Mode (Parallelism)

Image	Codec	Time (s)	Size (KB)	PSNR (dB)	FPS
Generic	libx264	20.58	26,116	41.36	280.00
Optimized	libx264	18.34	26,116	41.36	279.25
Generic	libx265	99.44	17,288	39.44	44.50
Optimized	libx265	77.57	17,440	39.36	56.00
Generic	libsvtav1	15.82	29,224	41.94	391.75
Optimized	libsvtav1	13.57	29,244	41.95	475.50

Chunked Mode Performance Improvement:

• libx264: 10.00% faster (Speedup: 2.24s)
• libx265: 21.00% faster (Speedup: 21.87s)
• libsvtav1: 14.00% faster (Speedup: 2.25s)

Optimizations applied

• Target CPU: -mcpu=neoverse-n1 (Optimized for the Ampere Neoverse-N1 architecture)
• Link-Time Optimization: -flto=auto enabled across all libraries and FFmpeg for improved inter-procedural optimization.
• Compiler Flags:
  • Standard optimization level (Default -O2) used for stability and peak performance on N1.
• Libraries: libx264, libx265, libvpx, libsvtav1
• Architecture: Built specifically for ARM64 / Ampere N1.

Credits

This project was created by Some Watson with the assistance of opencode, an AI software engineering agent.