Ask HN: How to Choose a Laptop for ML and Deep Learning?

3 points by aaossa ↗ HN
Hi HN,

I'm currently looking forward to join a master's program after my current career (in the lines of software engineering). I want to be ready to work with machine learning and deep learning libraries and be able to develop state of the art programs.

My problem is that I don't know what requirements should I be looking for in a laptop. What do you use? Do you have any laptop recommendation?

Everything is useful :)

Thanks!

9 comments

[ 2.6 ms ] story [ 38.3 ms ] thread
If it has to be a laptop, one along these lines: https://www.nvidia.com/en-us/geforce/products/10series/lapto...

You'll want a good GPU to train deep neural nets, and most libraries require one with CUDA support, i.e. an NVIDIA card.

The GTX 10 series essentially dropped the distinction between desktop and mobile cards, so you won't be giving up as much peak performance as in the past by getting a laptop. I still prefer to do such things on a desktop system with plenty of power and cooling headroom, where running at full tilts for hours or days on end is not a problem. Doing the same on a laptop feels like asking for trouble.

Thanks for your reply! Do you know if CUDNN support makes a difference in this case?
The list of CUDNN-supported frameworks,

https://developer.nvidia.com/deep-learning-frameworks

is pretty much a Who's Who of all the most popular ones. In that sense, CUDNN support does not make much of a difference - any framework you're likely to choose has it, and any recent NVIDIA card will run it.

A machine learning laptop doesn't make a lot of sense - you won't find a laptop with a comparable GPU. Really, the answer is whatever you are comfortable working on - you'll be connecting to a server or AWS GPU instance to actually train any large net and generally writing and testing on a laptop. That being said, get anything with an NVIDIA card (for the time being everything runs on CUDA) or that testing will be done on the CPU, though perhaps there will be libraries supporting OpenCL down the road.
> you won't find a laptop with a comparable GPU

Sure you will:

https://arstechnica.com/gadgets/2016/08/nvidia-pascal-laptop...

The real problem is sustained performance within a laptop form factor, as the article says:

How well the GTX 1080 maintains that speed within the confines of a laptop is largely going to depend on the design chops of laptop makers

It gets hot in those tight little boxes.

So if I want to use something like this I should be looking for a desktop pc?
I would definitely prefer a desktop for this kind of thing, yes.
Yes, you're right. I was thinking in using an NVIDIA card that allows me to use CUDA and CUDNN. Do you know if CUDNN makes a difference? Thanks!