Things to know
GPU are only available at specific zones.
GPUs are currently only supported with general-purpose N1 machine types.
Performance comparison of GPUs (T4, P4, V100, P100, K80)
The cheapest GPU available is NVIDIA® Tesla® T4 (16GB GPU memory), which cost $0.35/hour or $178.85/month after 30% sustained use dicount. If you don't to run it all the time, you will need to pay a higher rate (less sustained use discount).
For further discount, you can opt for preemptible (use the resource when no one else is using) or committed use discounts (pre-pay for resource reservation for 1 or 3 years). Check GPU pricing.
Google GPU is actually Google Compute Engine, which you can't run an instance with just GPU. You need a CPU, Memory and Disk, and charged for Network Usage as well. The cheapest configuration pricing is as of the following
1 vCPU with 3.75GB memory, attached with 1 NVIDIA Tesla T4 (16 GB memory)
$207.92 per month estimated
Effective hourly rate $0.285 (730 hours per month)
Item Estimated costs
Google Compute Engine Costs
VM instance: 1 vCPU + 3.75 GB memory (n1-standard-1) $34.67/month
Standard persistent disk: 100 GB $4.80/month
NVIDIA Tesla T4 GPU $255.50/month
Sustained-use discount − $87.05/month
Total $207.92/month
There are 2 ways to start a GPU instance
- https://cloud.google.com/ai-platform/deep-learning-vm/docs/tensorflow_start_instance (this is easier with NVIDIA/CUDA/TF drivers and libraries pre-setup)
- https://cloud.google.com/compute/docs/gpus/add-gpus#create-new-gpu-instance
Create GPU Instance with Deep Learning VM Marketplace
- Goto https://console.cloud.google.com/marketplace/details/click-to-deploy-images/deeplearning and click
Launch
- Deployment name:
test-gpu
- Zone:
us-west1-b
(check zones which support GPU) - Machine type: 1 vCPU (3.75 GB memory)
- GPUs
- Number of GPUs:
1
- GPU type:
NVIDIA Tesla T4
- Number of GPUs:
- Framework: TensorFlow Enterprise 2.3 (CUDA 11)
- GPU -
Check
Install NVIDIA GPU driver automatically on first startup? OPTIONAL
: Access to the Jupyter Lab - Enable access to JupyterLab via URL instead of SSH. (Beta)- Boot Disk
- Boot disk type:
Standard Persistent Disk
- Boot disk size in GB:
50 GB
(Although minumum is 30GB, but you bump intoRequested disk size cannot be smaller than the image size (50 GB)
error during instance creation).
- Boot disk type:
- Click
Deploy
.
You might see You've gone over GPUs (all regions) quota by 1 GPU. Please increase your quota in the quotas page.
at the top of the page, where you need to increase your GPU quota from 0 to 1.