News for hackers. Stuff that matters.
Re:Cost Per 1M tokens Of Self Hosting Llama-3 by causal on Friday June 14, 2024 @12:47PM (#40682391)
Came here to say this. No way you need to spend more than $1500 to run L3 8B at FP16. And you can get near-identical performance at Q8 for even less.

I'm guessing actual break-even time is less than half that, so maybe 2 years.

Re:Cost Per 1M tokens Of Self Hosting Llama-3 by causal on Friday June 14, 2024 @01:20PM (#40682716)
Furthermore, the AWS estimates are also really poorly done. Using EKS this way is really inefficient, and a better comparison would be AWS Bedrock Haiku which averages $0.75/M tokens: https://aws.amazon.com/bedrock/pricing/

This whole post makes OpenAI look like a better deal than it actually is.

Re:Cost Per 1M tokens Of Self Hosting Llama-3 by mrinterweb on Friday June 14, 2024 @01:36PM (#40682885)
I was getting that sense too. It would not be difficult to build a desktop machine with a 4090 for around $2500. I run Llama-3 8b on my 4090, and it runs well. Plus side is I can play games with the machine too :)