Advise regarding buying a GPU

User 1319 | 2/28/2015, 12:06:32 AM

I am very impressed by GraphLab ability to hand very large data sets. I was able to process ~100 GB worth of (~80,000) images on my 6 years old laptop (HP Pavilion Dv6, Intel® Core™2 Duo CPU P8700 @ 2.53GHz × 2, 4 GB RAM).

Now, I am looking to buy a laptop equipped with a powerful GPU to be able to apply deep learning to the processed dataset, and used the trained model to extract features to be used in building other models (similar to the Imagenet pretrained model approach).

As you know, their are many NVIVIDA GPU models http://www.nvidia.ca/object/wtb_notebooks.html, and I am struggling with choosing a laptop with a GPU suitable for GraphLab deep learning modelling.

I understand that you used NVIDIA GTX-780 GPU with the MNIST image dataset. Did you use the same GPU with the Imagenet deep learning model?

Could you please advise me which type of GPUs should I consider.

For example, I am looking at the Dell Precision M3800 Workstation which comes with NVIDIA® Quadro K1100M graphics with 2GB3 memory. Does the Quadro K1100M provide good performance with deep learning? http://www.dell.com/ca/business/p/precision-m3800-workstation/pd

Thank you in advance. Cheers

Tarek

Comments

User 1190 | 3/2/2015, 6:24:16 PM

Hi,

Yes we use GTX-780 GPU for training the imagenet model. When choosing GPUs there are three main things to consider - Number of processors - Amount of device memory - Memory bandwidth and interface (128bit or 256bit or higher)

The GTX780 has 2304 cores, 3G memory, and 288G/sec memory bandwidth (384bit) http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-780/specifications

Quadro K1100M has 384 cores, 2G memory, and 44G/sec memory bandwidth (128bit) http://www.nvidia.com/object/quadro-for-mobile-workstations.html

Whether does the card you are looking at provide good performance with deep learning depend on your problem.

For your reference, using GTX780, the throughput for imagnet is about 200 images / sec, or 100 mins per iteration. Quadro K1100M is likely to be at least 7 times slower.

Hope this can help you choose your hardware for GraphLab Create and deeplearning.

-jay


User 1319 | 3/2/2015, 9:32:00 PM

Hi Jay, Thank you very much for your detailed reply. Highly appreciated. Tarek