Cost effectiveness of the different Cloud/Server renting options


Worked like a charm!!
Thanks a lot @aseem :slight_smile:

(Antoine) #24

Hello everyone,

First of all, thanks a lot for this awesome course material, I’m really glad I found it !

I’m new here (currently watching lesson 2) and I was wondering what is the total cost of using paperspace by the end of lesson 7 (considering the toying with the code)?

I’m a bit reticent about paid plans since it’s only for experimentation purposes at the time but if the cost is moderate, I might switch to a cloud option to train models faster (and spend less time waiting for computations).

(Max) #25

Hi @malrod,

How are you running your jupyter when you don’t need the GPU computation power on google cloud?

(Vaisakh) #26

You can add/remove GPUs, change from one GPU type to another,
and increase/decrease the number of GPUs.

So, for example, if I’m just starting off and want to download and preprocess datasets, I’ll turn the specs down to an f1-micro (~$0.004/hr) or an n1-standard-2 (~$0.021/hr).

Then, if I need a GPU for quick prototyping, I’ll add a K80 (~$0.243/hr).

For small projects, I’ll have a P100 n1-standard-8 instance (~$0.819/hr) for training.

For large projects, a V100 (~$1.261/hr).

If it’s a really large project that’ll take lots of time otherwise, and I see that multi-GPU training helps,
I’ll take 8 of those, thank you :blush:

Note: The prices are for preemptible instances.

(Max) #27

Awesome idea. Thanks will do just that! Downloading and storing data sets on google cloud is relatively inexpensive too I suppose?

(Vaisakh) #28

It’s $0.170/GB per month of persistent SSD storage.