TIL

Picking cloud GPU provider for experiments

Looking for cheap one for few hours

Google Cloud https://cloud.google.com/compute/gpus-pricing

AWS https://instances.vantage.sh/?region=eu-west-1

Lambda labs https://lambdalabs.com/service/gpu-cloud

For today prices per hour for

1xNVIDIA A10 with 24GB VRAM

Lambda Labs:

• 1x NVIDIA A10 24 GB 30 200 GiB 1.4 TiB $0.60 / hr

GCloud

• not in offer

AWS

• g5.2xlarge 32.0 GiB 8 vCPUs 1 NVIDIA A10G 24 GiB 450 GB NVMe SSD Up to 10 Gigabit $1.3530 hourly

1xNVIDIA A100 with 40 GB VRAM

Lambda Labs:

• 1x NVIDIA A100 40 GB 30 200 GiB 512 GiB $1.10 / hr

GCloud

• NVIDIA A100 40GB 1 GPU 40 GB HBM2 $2.934 per GPU

AWS

• p4d.24xlarge 1152.0 GiB 96 vCPUs 8 NVIDIA A100 320 GiB 8000 GB (8 * 1000 GB NVMe SSD) 4x 100 Gigabit $35.3965 hourly

I have to figure out how to use this instance only for necessary training time, and automate its shutdown to reduce costs:

Lambda Labs instaces does not support instance shutdown, only termination, so first the data must be stored to some block storage.

https://docs.lambdalabs.com/cloud/terminate-instance-api/?_gl=11od6ru3_gaOTg5NjgxMzY1LjE2ODEyMzE0NDQ._ga_43EZT1FM6Q*MTY4MTIzMTQ0NC4xLjEuMTY4MTIzMjE0Ni40Mi4wLjA.

AWS, Gcloud supports proper instance shutdown