for a TPU with more than 16GB GRAM and less than 40GB, so sometime we need to provision a A100 to get the training speed we want but we don't use all the GRAM

Oh that makes sense...
Just saw this one, this might help?
https://www.globenewswire.com/news-release/2022/10/24/2539924/0/en/ClearML-and-Genesis-Cloud-Announce-New-MLOps-Partnership-Delivering-100-Green-Energy-Compute-Solution-for-Machine-Learning.html

I think it's supposed to be out early Nov 🙂

Hi FierceHamster54
This is already supported, unfortunately the open-source version only supports static allocation (i.e you can spin multiple agents and connect each one to specific number of GPUs), the dynamic option (where you have single agent allocating jobs to multiple GPUs / Slices is only part of the enterprise edition
(there is the hidden assumption there that if you spent so much on a DGX you are probably not a small team 🙂 )

Hey, I'm a SaaS user in PRO tier and I was wondering if it was a feature available on the auto-scaler apps so I could improve the cost-efficiency of my provisionned GCP A100 instances

There is a gap in the GPU offer on GCP and there is no modern middle-ground for a TPU with more than 16GB GRAM and less than 40GB, so sometime we need to provision a A100 to get the training speed we want but we don't use all the GRAM so I figured out if we could batch 2 training tasks on the same A100 instance we would still be on the winning side in term of CUDA cores and getting the most of the GPU-time we're paying.

I could improve the cost-efficiency of my provisionned GCP A100 instances

But their pricing is linear, if you do not need a100 get a cheaper instance ?! no?

Oh wow, would definitely try it out if there were an Autoscaler App integrating it with ClearML