search

ovh / public-cloud-roadmap

Agile roadmap for OVHcloud Public Cloud services. Discover the features our product teams are working on, comment and influence our backlog.
https://www.ovhcloud.com/en/public-cloud/
186 stars 5 forks source link

GPU L4 Flavors in GRA11 #502

Closed JacquesMrz closed 7 months ago

JacquesMrz commented 1 year ago

Summary

NVIDIA L4 GPUs with 24GB of memory. The L4, based on the NVIDIA Ada Lovelace GPU architecture, is a universal GPU for every workload with enhanced AI and video capabilities. It provides efficient compute resources for graphics, simulation, data science and data analytics.

Intended Outcome

L4 range of instances, flavors and pricing details to come. Flavor Name RAM vCore GPU Local Storage Public Network Private Network
L4-90 90 GB 22 L4 24GB x 1 400 GB NVMe 8 Gbps 8 Gbps max.
L4-180 180 GB 45 L4 24GB x 2 400 GB NVMe 16 Gbps 16 Gbps max.
L4-360 360 GB 90 L4 24GB x 4 400 GB NVMe 25 Gbps 25 Gbps max.

How will it work

Standard way to start a GPU instance, selecting the right L4 flavor. https://help.ovhcloud.com/csm/en-public-cloud-compute-deploy-gpu-instance?id=kb_article_view&sysparm_article=KB0050735

JacquesMrz commented 7 months ago

We removed the allocation of NVMe disks in Passthrough because of the constraints it was inducing regarding migration options from one host to another.

bboigienman commented 6 months ago

do you have plan to make L4 available in other GRA zones or regions?

JacquesMrz commented 6 months ago

Hi @bboigienman , can I ask you where exactly you would this L4 flavors available ? In addition, is there any specific reason why GRA11 is not ok for you ?

bboigienman commented 6 months ago

Hi @JacquesMrz, our use case is High Availability setup for production inference (millions of requests per day). This is the reason relying on a single datacenter is not compatible with our prod env requirements. In terms of location, it would be better in another French region (Roubaix) or in Gravelines (as far as possible from GRA11). Thank you