Laptops for data science, deep learning, & machine learning

[rockfire.sachin]

Any suggestions which laptop to use for machine learning and data science at affordable price.

[AskTJ]

Budget and country. Always list these.

[Krolic]

I currently do research in machine learning and deep learning (image segmentation research), and I can tell you there isn't much out there for laptops. The main thing with deep learning is GPU acceleration, so the only type of laptop would be one of those massive gaming monstrosities that are basically desktops in "portable" form factors. And even with one of those, you couldn't do a lot with it, since a model could take hours to train depending on how deep it is or what not. And this doesn't take into account thermal throttling, since the GPU would be going non-stop for hours. You also would not be able to use the laptop for anything while it is training. Unfortunately, there is not much availability out there. That being said, depending on where you are and the work you are doing, you may be able to find servers to connect with. With my work I am able to connect to universities clusters to do computing off of, which have significantly more power than any laptop available. 

[RimsOnAToaster]

So maybe I could help here. Like @Krolic said, GPU acceleration is the key to any type of machine learning. I've been working in the recurrent neural network space (speaker recognition for animals) and because my datasets are relatively small in scope, I've been able to get away with using a 2014 13" MacBook Pro connected via Thunderbolt 2 to an external graphics card (or eGPU) for some time now. Basically, Thunderbolt is a connector type that can carry PCIe signals over a wire, which means that with a combination of a Thunderbolt-enabled laptop, a Thunderbolt-enabled enclosure, and a graphics card, you can get the benefits of a desktop graphics card without it being built into your laptop.

 

I recommend using any laptop with thunderbolt 3, a quad core processor, and whatever you can find for an enclosure and graphics card.  Check out this website for more info: egpu.io

[Krolic]

31 minutes ago, RimsOnAToaster said:

So maybe I could help here. Like @Krolic said, GPU acceleration is the key to any type of machine learning. I've been working in the recurrent neural network space (speaker recognition for animals) and because my datasets are relatively small in scope, I've been able to get away with using a 2014 13" MacBook Pro connected via Thunderbolt 2 to an external graphics card (or eGPU) for some time now. Basically, Thunderbolt is a connector type that can carry PCIe signals over a wire, which means that with a combination of a Thunderbolt-enabled laptop, a Thunderbolt-enabled enclosure, and a graphics card, you can get the benefits of a desktop graphics card without it being built into your laptop.

 

I recommend using any laptop with thunderbolt 3, a quad core processor, and whatever you can find for an enclosure and graphics card.  Check out this website for more info: egpu.io

Only needing a single GPU would be great. My stuff is with 3D images so I generally just grab 3 GeForce Titan X's off of our cluster and it still takes hours.

 

@RimsOnAToaster is right, getting an external GPU would work, but with a laptop with Thunderbolt and the eGPU could get expensive fast. 

[RimsOnAToaster]

53 minutes ago, Krolic said:

My stuff is with 3D images so I generally just grab 3 GeForce Titan X's off of our cluster and it still takes hours.

Oh god, that sounds way out of my league. I upgraded from a GTX 750 Ti when I was first learning to a GTX 970 at the beginning of the project and thought that was the end-all-be-all. Right now I use a GTX 1070 in a clean, warranty-friendly enclosure from Gigabyte and I can crunch through about a days-worth of recordings over a long lunch. Pascal is truly amazing - I can't wait to see what Volta has in store.

 

 

My eGPU was pretty competitively priced at $600 for the graphics card and enclosure. I think another $900 for a base-model laptop with Thunderbolt 3 sounds about right? So all in it would be around $1500. That's expensive for sure, but not impossibly so.

[TLCH723]

You probably just use the laptop just to prototype some of the code and then send it to a workstation or better yet HPC.

 

There are two options, a laptop with eGPU or a workstation

[VanBantam]

A bit late to the party here so I guess more for posterity than helping the OP.

 

I would caution against getting an external gpu and instead use a cloud service like paperspace.  For the amount of money you'll be spending on the beefy laptop you can get a good chunk of time with a modern server grade gpu.  For the amount of money it would cost to buy an external gpu you can build a small form factor pc with a 1080 ti, 2080 ti, or titan V and not deal with the added layer of complexity and bottleneck that thunderbolt will cause.