As a web developer, I am growing increasingly interested in data science/machine learning, enough that I have decided to build a lab at home.

I have discovered the Quadro RTX 4000, and am wondering how well it would run ML frameworks on Ubuntu Linux. Are the correct drivers available on Linux so that this card can take advantage of ML frameworks?


This is the only driver that I could find, but it is a "Display Driver", so I am not sure if that enables ML frameworks to use this GPU for acceleration. Will it work for Intel based processors?

Any guidance would be greatly appreciated.

  • $\begingroup$ I've used 2000 version and the major point is that it does not have a good memroy. $5GB$ is not appropriate for DL tasks. If you can afford it, buy a 2080 which is perfect. I don't know the memory of 4000 but the 2000's memory is very limiting and you cannot train big models on it. But the gpu itself is roughly a powerful one. $\endgroup$ Apr 5 '19 at 21:33
  • $\begingroup$ I can also refer that PNY does not have a good cooling system. You have to take that in mind. $\endgroup$ Apr 5 '19 at 21:37
  • $\begingroup$ Thanks for your feedback @Media. Would you be able to recommend a card that would work well for getting up and running with ML/Deep learning? $\endgroup$
    – crayden
    Apr 5 '19 at 21:38
  • $\begingroup$ I guess 2080ti is the best at the moment due to its power and new tensor modules that have been introduced inside them for DL/ML tasks. It is also far cheaper than titan. $\endgroup$ Apr 5 '19 at 21:40
  • $\begingroup$ I noticed you are referring to the former 2000/5GB version of the Quadro. The new Quadro RTX line is based on the Turing architecture, and includes special tensor cores for acceleration. This should make a huge difference between the 2000 version you have used, and the new RTX/Turning based cards? $\endgroup$
    – crayden
    Apr 5 '19 at 21:45

You seem to be looking at the latest Quatro 4000, which has the following compute rating:

enter image description here

You can find the complete list here for all Nvidia GPUs.

While it seems to have an impressive score of 7.5 (the same as the RTX 20180ti), the main draw back the memory of 8Gb. This is definitely enough to get started with ML/DL and will allow you to do many things. However, memory is often the thing that will slow you down and limit your models.

The reason is that a large model will require large number of parameters. Take a look at the following table (models included in Keras), where you can see the number of parameters each model requires:

enter image description here

The issue is that the more parameters you have, the more memory you need and so the smaller the batch size you are able to use during training. There are many arguments for larger vs. smaller batch sizes - but having less memory will force you to still to smaller batch sizes when using large models.

It seems from Nvidia's marketing, that the Quadro product line is more aimed towards creative developers (films/image editing etc.), whereas the Geforce collection is for gaming an AI. This highlights that Quadro is not necessarily optimised for fast computation.

  • $\begingroup$ Nvidia's marketing makes be believe this card is better than GeForce for AI? From the product page: "Deep learning frameworks such as Caffe2, MXNet, CNTK, TensorFlow, and others deliver dramatically faster training times and higher multi-node training performance. GPU accelerated libraries such as cuDNN, cuBLAS, and TensorRT deliver higher performance for both deep learning inference and High Performance Computing (HPC) applications." I thought GeForce was optimized for gaming, in contrast. $\endgroup$
    – crayden
    Apr 5 '19 at 23:08
  • $\begingroup$ Perhaps they are starting to push in that direction - they have the same text everywhere, but Geforce has generally been the product line to go for. You care about number of cuda cores, amount of memory and the transfer rate of that memory. Then just find the best combination of those factors that your budget allows. $\endgroup$
    – n1k31t4
    Apr 5 '19 at 23:20
  • $\begingroup$ What is an adequate amount of memory and CUDA cores? $\endgroup$
    – crayden
    Apr 5 '19 at 23:24
  • $\begingroup$ How long is a piece of string? ;) if you want to work with images/videos, the more the better. Working with text can be less memory intensive and something like stock market data is not memory hungry. If you get the Quadro, an RTX or a Titan - it is likely that the human will be the slowest link. Just don't work with a CPU and you'll be fine. $\endgroup$
    – n1k31t4
    Apr 5 '19 at 23:32

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.