Server Configuration

I have the very fortunate opportunity to have my personally configured server. I asked the Professor of my University if I could get access to the University’s computation power to dive deeper into Deep Learning (which was part of our curriculum).

He agreed and I now have to send him some details on how I want the server to be configured and wanted to ask If you could give me some advice on what to choose, as I have little experience.

Could you give me some numbers on the following?

  • computer memory
  • number of logical processors
  • disk storage

The offering of the professor was very generous so I don’t want to go overboard with my demands.

-Fabian

ps: There are no dedicated machine learning GPUs available, but with enough computational power the course should be doable without, right?

I would probably request:
4 vCPU
32 GB RAM
100 GB Storage

The P2 instance on AWS that Jeremy uses has:

1 GPU (12 GB)
4 vCPU
61 GB RAM

I think you can compare what you will need to a t2.large instance, but I haven’t went through part 1 on the AWS instance so I’m not 100% sure.

That one has:
0 GPU
2 vCPU
8 GB RAM

The specs of the server are next to irrelevant to be honest. You could try going for as strong of a CPU as you can get but that is about it.

There are many cool things you can do with a CPU only system but I do not think you will be able to complete the assignments of this course without access to a GPU, any GPU for that matter.

I am not even sure why I put this together a couple of months ago when I was working through part 1 of the course for the first time, but this is something you could work through if you’d like which you could do on a CPU and is roughly equivalent to the cats vs dogs that the first couple of lessons focus on.

Jupyter notebook with worked example
Boilerplate to get you started on your own hacking away

The MNIST database is quite fun IMHO and you could add the whole data augmentation and all the other techniques that Jeremy describes.

I’d recommend having a good read through this excellent thread. Making your own server

It’s long. But you’ll learn a lot.

You could offer to buy your own GPU and have it installed in the University’s computer. Even a second-hand / older model GPU is better than no GPU.

He would be able to do the course without a GPU wouldn’t he? It’s just going to be slow to fit the model. Let me know if I’m wrong on that and there is a technical limitation. I agree he would want a GPU if possible, but if that’s all the school has, he could still work through the course and maybe even show them why a server with some GPUs would be a worthwhile investment.

Thanks for the great answers so far!

I was wondering the same thing as @KevinB, wouldn’t I be able to do the course with just CPUs? Besides the time it would take, what would make it impossible?

I do understand the necessity of GPUs, specifically for Machine Learning and I will suggest it at my university :slight_smile: