Machine Learning System Configuration

[Jeetendra Vaghela]

Hi All,

I am working on machine learning project where I used "500K" images to train the model. I am using lib-linear SVM(Support Vector Machine) With HOG (Feature Descriptor). HOG take 128X 128 images as input and produces 1096 features. Produced features vector (1096 X 500K) used as input to SVM Model Training.

My Current System Configuration is:

• Processor : Intel® Core™ i7-6700 CPU @ 3.40GHz × 8
• RAM: 16 GB
• Graphics: Intel® HD Graphics 530 (Skylake GT2) 

which took around 5 hours and 30 minutes to train the model.

 

I also tested the same task on AWS cloud and the results are:

Instance Type    CORE    RAM    Training Time In (Hours)            
c4.8xlarge          36          60         03:39:15
c5.4xlarge          16          32         01:23:22
 

Now, I want to build my own system and I am really confused to select processor.

Should I go for Intel® Xeon® Processor E5-2699 v4 with 22 cores OR INTEL® CORE™ i9-7980XE EXTREME EDITION PROCESSOR with 18 cores?

 

Your help and suggestion would really be appreciated.

Thank you.

 

[Guest]

What are you using? AFAIK gpus are far better at machine learning related workloads especially in tensorflow. I'd personally go for the Xeon considering it has double the cache and ~22% more cores, but I can't make a definite decision there as I don't know what you are using specifically and I don't do much in this field.

Edit: speaking of tensorflow, the Titan V is a beast there (Has a good number of tensor cores)

[Jeetendra Vaghela]

I am using lib-linear SVM(Support Vector Machine) With HOG (Feature Descriptor).

HOG take 128X 128 images as input and produces 1096 features. Produced features vector (1096 X 500K) used as input to SVM Model Training.

[Jeetendra Vaghela]

36 minutes ago, tjcater said:

What are you using? AFAIK gpus are far better at machine learning related workloads especially in tensorflow. I'd personally go for the Xeon considering it has double the cache and ~22% more cores, but I can't make a definite decision there as I don't know what you are using specifically and I don't do much in this field.

Edit: speaking of tensorflow, the Titan V is a beast there (Has a good number of tensor cores)

Thank you for your reply.

[Guest]

10 minutes ago, Jeetendra Vaghela said:

I am using lib-linear SVM(Support Vector Machine) With HOG (Feature Descriptor).

After doing some brief research on what this is, I can confirm that this goes way beyond my minuscule understanding of the field.  

12 minutes ago, Jeetendra Vaghela said:

Thank you for your reply.

Np, sorry that I couldn't have provided more help. If this is highly parallelized like most ML I've read about, I'd suggest researching the prospects of GPU acceleration if you haven't already. Hopefully someone with better knowledge passes by and gives a more definitive answer.