VIP regarding CUDA Deep Learning accessibility on Laptops

I need to be able to run NVIDIA’s Deep Learning and Related Libraries (prototyping) on a Laptop
NOT a Desktop or Workstation.

I’m looking to use either GTX 980 or 980M.

Unfortunately no laptop vendor offers anything other than Windows.

I need to run Ubuntu 14.04.3 LTS for all the Deep Learning ML libraries to work.
Is this still true?

I have communicated with 1 researcher who has older GPUs running Ubuntu on Laptop but he said it’s not stable.

QUESTION: What Laptop options do I have?


PS- NVIDIA’s suggestion

Hi Greg,

This is a great question for the GPU Computing developer forums @

If you ask your question there and send me a link to it, I’ll make sure someone from NVIDIA checks it out.



PS- This may be the most important question the NVIDIA MACHINE LEARNING CUDA DEVELOPER GROUP needs to Solve ASAP.

The principal DL focused library offered by NVIDIA is cuDNN. There is a landing page for cuDNN that can be found by googling “cuDNN”:

where it states:

“Supported on Windows, Linux and MacOS systems with Kepler, Maxwell or Tegra K1 GPUs.”

You might want to be more specific, if you have questions about other specific software packages, like Caffe, Torch, Theano, etc. none of which are written by or maintained by NVIDIA. I believe Caffe and Theano can be configured on windows, for example:

If you want to run DIGITS:

AFAIK that is linux only.

I have no first-hand experience with Linux on laptops, and it is not clear what country you reside in, but typing “linux pre-installed on laptop” into the Google search engine results in a number of links for vendors selling laptops that come with Linux.

  1. Njuffa, obviously there are numerous Laptops that run Ubuntu and other Linux forms, I have one of them myself. However, that is NOT the problem!

The problem is that a Laptop is needed with a GTX 980M or higher GPU on-board that can run Ubuntu 14.04.3 to run DIGITS as well as cuDNN, cuBLAS, cuFFT,…

  1. Thanks TxBob. It seems that you believe I could run cuDNN without problem on Windows. You may be correct.I read that too but online I’ve seen people claiming it didn’t work properly. In addition, it can’t be used for an open source project if its dependent upon windows and visual studio, at least that’s what a cancer researcher told me.

That’s why an NV engineer with real knowledge needs to clarify my original question.

My apologies if I misinterpreted “Unfortunately no laptop vendor offers anything other than Windows.” Typing your configuration requirements into Google brings back links like this one
I am running Ubuntu 14.04 on a Sager NP8652 / Clevo P650SG with a Nvidia 980M graphics card.

which seems to indicate that laptops with the desired configuration exist (although it is not immediately clear whether they come pre-installed with Ubuntu 14.04, or some other Linux distro). You might want to follow-up the search engine leads. It seems that various vendors give their Linux customers a choice of distro.

Above, txbob already pointed to the official NVIDIA information for cuDNN that states that cuDNN is supported on Linux, as long as the GPU used is in the Kepler or Maxwell family.

Not all the options below but, assuming one uses Ubuntu 14.04.3, here are some options I have used with success for DL related workloads w/ cuDNN v3/v4:

  • MSI GT72 (w/ 980M):
    Pros: very stable. Does not require Optimus.
    Cons: need a 4.x+ linux kernel including the ath10k module for wifi (latest 4.2 avail from Ubuntu works). Large format (a lighter laptop might be best if traveling).
  • Gygabyte P35X v3 (w/ 980M):
    Pros: Works out of the box. Slim. Affordable.
    Cons: unstable linux touchpad support (the pointer would randomly disappear sometimes by default).
  • Razer blade 2015 (w/ 970M) (also tested on Debian stretch)
    Pros: well built, silent, slim and great for travel. Great screen resolution.
    Cons: backlit keyboard LEDs cannot be easily switched off. ACPI not fully supported. Integrated USB camera not fully uvc supported (works in some apps but not all by default). Multi-touch support requires 4.x kernel or patch.

In any case, Ubuntu does not come preinstalled on the aforementioned laptops. 14.04.3 heavily recommended. Results above are reported from a EFI based install.

@JulieBernauer could you please provide more details, or better - step-by-step instructions, on how you managed to successfully install nvidia driver and cuda on the MSI GT72?

My laptop model is GT72-2QE Dominator Pro (w/ 980M), to which I added a second SSD (nonRAID mode) and installed Ubuntu 14.04.3 alongside Windows 8.1.

My preference would be to continue using Ubuntu since I only just got accustomed to it from a lifetime of Windows, but if necessary I am willing to make compromise on this behalf. I have been struggling with setting up this environment for the past several days and didn’t manage to make it work. Tried many online tuts, with both the .deb and .run for cuda 6.5 toolkit.
The desire is to use the setup for dev in CV and DNN, along with my Jetson TK1 kit.

The issues that I run into are missing dependencies (cuda &nvidia related, that point to libcheese & others and xorg-video-abi-xx). Also if I get pass these, the issues are black screen or login loop.

Thank you in advance for any help you can provide.

For your installing cuda and the driver on MSI GT72:

Go to:

Choose: Linux-> x86-64 -> Ubuntu -> 14.04 -> deb(network)
and follow the instructions, you’ll have cuda and the driver installed that way and on a clean Ubuntu installs this works fine.

Note that you want cuda 7.5 since it comes with a 352 driver (you want a driver which is at least 346.35).
If having issues with display, one wants to check nvidia-prime/nouveau is not enabled.

For Jetson, I would recommend using jetpack which will take care of adding the right dependencies for the right architecture (since Jetson is armhf) on the host machine and push the required dependencies on the board. Cheese and all then work fine.

Has anyone been able to use a MSI GT72 with the 970M and a linux distro. I note here this works with the 980M variant. Will it be any different if I used a 970M. For my development activity the 970M will provide adequate processing Cuda cores.

Have read of display issues with the MSI GT72 but nothing in recent posts on if it was resolved.