• 0 Posts
  • 13 Comments
Joined 4 months ago
cake
Cake day: March 4th, 2024

help-circle


  • The A16 Bionic has as Neural Engine capable of 17 TOPS but 6GB of RAM.

    The M1 had a Neural Engine capable of just 11 TOPS but all M1 chips have at least 8GB of RAM.

    So the model could run on an A16 Bionic if it had 8GB of RAM as it has 54% more TOPS than the M1, but it only has 6GB of RAM. Apple have clearly decided that a model small enough to fit just wouldn’t give good enough results.

    Maybe as research progresses they’ll find a way to make it work with a model with fewer parameters but I’m not going to hold my breath.







  • The thing with serverless is you’re paying for iowait. In a regular server, like an EC2 or Fargate instance, when one thread is waiting for a reply from a disk or network operation the server can do something else. With serverless you only have one thread so you’re paying for this time even though it’s not actually using any CPU.

    While you’re paying for that time you can bet that CPU thread is busy servicing some other customer and also charging them.

    I like serverless for it’s general reliability, it’s one less thing to worry about, and it is cheap when you start out thanks to generous free tiers, at scale it’s a more complex answer as whether it is good value or not.