Faster FFmpeg on Graviton3

Overview

I was recently investigating the performance of FFmpeg on Intel instances, and was curious how much faster ARM instances could perform after reading this blog post from AWS. At the time of this writing, AWS Graviton3 claimed to offer the best price to performance in EC2 for compute-intensive workloads, and was also one of their relatively newer offerings.

To take advantage of these improvements, you will need to be using v6.0 or newer of FFmpeg. The base instance that I compared this ARM instance to was a 3rd Generation Intel Xeon. I also utilized GNU parallel to efficiently pipeline some of my commands.

Benchmark

First, I needed to extract the frames from my source piece of media on each machine:

#!/bin/bash
ffmpeg \
	-i 'output.mjpeg' \
  -c copy \
  -map 0  \
  -segment_time 00:00:03 \
  -f segment frames/output%03d.mjpeg

Finally, I applied a quick colorbalance filter in parallel to our .jpeg frames:

#!/bin/bash
time parallel 'ffmpeg -i {} -vf colorbalance=gm=0.25 -y {}.mp4' ::: frames/*.mjpeg

Results

Honestly, I was expecting about a 20-30% speed boost. The c6i took 9.4 seconds averaging 10 runs, and the c7g took 3.9 seconds averaging 10 runs, for a ~240% performance boost. Not bad at all!