6 ms·

Some anecdotal data, but we recently estimated the cost of running a LLM at $WORK by looking at power usage over a bursty period of requests from our internal u

Some anecdotal data, but we recently estimated the cost of running a LLM at $WORK by looking at power usage over a bursty period of requests from our internal users and it was on the order of $10s/mil tokens. And we arent a big place, nor were our servers at max load, so I can see the cost being much lower at scale

exceptione 1y ago
This is only the power usage?
- dist-epoch 1y ago
  Hardware spend also need to be amortized (over 1 year? 2 years?) Unless you cloud rent.
  - theOGognf 1y ago
    That is true too
  - jenny91 1y ago
    5 year amortization is pretty realistic I'd say. A100s (came out 2020Q1) are still in heavy use. (I think V100s from 2017Q3 are starting to be phased out a fair bit.)
- theOGognf 1y ago
  Right, this is only power usage. Factoring in labor and all that would make it more expensive for sure. However, it’s not like it’s a complex system to maintain. We use a popular inference server and just run it with some modest rate limits . It’s been hands-off for close to a year at this point
  - exceptione 1y ago
    Ok! What hardware do you run? I had thought that would be the most expensive part.