This stack makes it easy to compare Graviton2 (ARM64) and Intel (X68_64/AMD64) architecture performance on AWS Lambda for CPU-intensive video file conversions using ffMpeg.

The stack sets up a S3 bucket and two Lambda functions. Both functions take any file uploaded to the S3 bucket, resample it using ffMpeg, and output the time for the ffMpeg job (ignoring lambda startup and s3 file download time).

• GNU make or compatible makefile runner
• aws command line tools (optionally AWS SAM to look for logs easily)
• Node.js 14 or later

Deploy the stack to your aws account (replace DEPLOYMENT_BUCKET value with your bucket, optionally setting the max allowed memory size as well):

make deploy MEMORY_SIZE=10240 DEPLOYMENT_BUCKET=my-cf-deployment-bucket

by default the stack name will be graviton-test, you can change that by setting STACK_NAME during deployment, similar to the parameters above.

Get the upload bucket name for the stack

make outputs

Upload a video to the bucket for conversion (get the upload bucket name from the previous command)

aws s3 cp video-file.mp4 s3://UPLOAD_BUCKET_NAME

Look for logs in cloudwatch, or using AWS SAM:

sam logs -n ArmConvertFileFunction --stack-name graviton-test
sam logs -n IntelConvertFileFunction --stack-name graviton-test

Each function logs to its own cloudwatch logs, with a JSON format outputing the uploaded key, max available memory size for lambda

• Graviton 2/ arm function

{
  key: 'pexels-i-am-sorin-6172942.mp4',
  memorySize: '10240',
  arch: 'arm64',
  time: 53565,
  fileSize: 184974533
}

• Intel/X86 function

{
  key: 'pexels-i-am-sorin-6172942.mp4',
  memorySize: '10240',
  arch: 'x86_64',
  time: 43454,
  fileSize: 184974533
}

• 1 October 2021, us-east-1 region: https://www.narakeet.com/tech/graviton-lambda-ffmpeg.html (Intel wins by more than 20% in average)