First draft of methodology for modeling generative AI systems #221
Conversation
| @@ -0,0 +1,12 @@ | |||
| ``` | |||
There was a problem hiding this comment.
where is this snippet used?
There was a problem hiding this comment.
It's not, but figured we'd want it when we do the model later? Wasn't sure if we'd do an AI variant of calculations.mdx, so really just here as a placeholder.
| | --------- | -------------- | | ||
| | Total reserved time | 118 days | | ||
| | Reservation start time | January 2022 (?) | | ||
| | GPU hours for final model | 1,082,990 | |
There was a problem hiding this comment.
is GPU hours suppose to be the hours where the GPUs were used during the intermediate and final training?
In this case with a cluster of 384 GPUs, this total GPU hours represents 117.5 (1082990/384/24) cluster days, which is the same as total reserved time.
I would have thought total reserved time in days to be more than the GPU hours for final model.
There was a problem hiding this comment.
Right - that's where we backfill the intermediate down below when we normalize. If you have ideas on how to clarify please lmk
|
|
||
| Embodied emissions = (cluster embodied emissions per hour) x (training time) | ||
|
|
||
| Usage emissions = (usage energy per GPU-hour) x (total GPU hours) x (average grid intensity during training) |
There was a problem hiding this comment.
if instead of total GPU hours, we had GPU hours utilized by hour/time, we could potentially leverage the actual grid mix during that hour. i'm thinking of cases where potentially training could be paused and resumed to only happen during times where marginal grid mix intensity is low
There was a problem hiding this comment.
Yep agreed - though I was thinking this would be easiest to represent as a lower average grid mix?
LR updating phase description. Testing for update process going forward
Remove extra word
No description provided.