Help me choose parts for my new PC....

[daanav]

Hey everyone,

I am still in planning phase before I order parts for my new PC.

I will be using this pc for mostly machine learning research hence my will be working mostly on Linux (Ubuntu Distro). Also I will be gaming on it a little bit.

I am very confused and not able to decide which cpu to get. The decision has been made on the following config, 

• Graphics Card : Asus ROG Strix GTX 1080 Ti
• 2 Tb Hard Disk 7200 rpm
• 500 Tb SSD
• 512 Gb Samsung 970 Pro M.2
• 32 Gb Ram

I am confused as to which CPU to get. I will be adding another GPU as my models and data get bigger, but will not exceed two GPUs.

I don't want to have bottlenecking because of the number of pcie lanes available with the cpu and also don't want to pay through the nose for a i9 processor. 

I don't know much about pcie lanes and how many do I need. I just know that I need to run the GPUs at 16x. If someone could explain that also, it would be great.

 

Does anyone have suggestions or advice regarding what cpu to get or even if this config is ok for the task of machine learning.

 

Thanks in advance.

[Firewrath9]

Id get a TR 1920X or 2920X. It has all the pcie lanes you will ever need. Else a 2700X or a i9-7900X.

[Quadrum]

AMD 2990wx

[daanav]

Thanks for suggesting AMD, I had also considered AMD but as I will be using Linux, there could be compatibility issues with Linux kernel.

[mymande]

if you can afford it the Titan V is made for machine learning

if you want something cheaper the amd vega x64  for example https://www.gigabyte.com/Graphics-Card/GV-RXVEGA64GAMING-OC-8GD#kf has better performance and faster memory (i do not know the situation of amd drivers for games in linux), I would also consider waiting for tomorrow Nvidia launch of the 2000 series.

threadripper is perfect although the i9 has better gaming performance, instead of going with one m. 2 and a normal ssd i would suggest 2 m.2 in raid 0 for better performance or if you are working with sensible data raid 1 for redundancy

[daanav]

@mymande Thanks for replying. I know Titan V is made for machine learning, but 1080 Ti can also be used for it, as there is not much difference between the two for machine learning use. 

Gaming for me is secondary, or a nice to have. This machine will 99.99% of the times be used for machine learning. 

[mymande]

10 minutes ago, daanav said:

 Thanks for replying. I know Titan V is made for machine learning, but 1080 Ti can also be used for it, as there is not much difference between the two for machine learning use. 

that is not quite true the titan v is based on the volta architecture created for the machine learning while the 1080 ti is pascal, of course you could use the 1080 and it will be fine but the vega one has a faster memory,

[daanav]

2 minutes ago, mymande said:

that is not quite true the titan v is based on the volta architecture created for the machine learning while the 1080 ti is pascal, of course you could use the 1080 and it will be fine but the vega one has a faster memory,

oh sorry, you are right, I somehow read it as Titan X. 

But I am more concerned about the CPU currently. 

[mymande]

33 minutes ago, mymande said:

threadripper is perfect although the i9 has better gaming

Then this will sum it up

If u do not want an i9 but still stick with Intel checkout the Xeon workstation line and see which one it is the better but I do not know if it will be cheaper.

If you need multicore performance you should checkout server stuffs

 

[daanav]

Hey guys,

 

How about i7-6850k. It offers 40pcie lanes. And I read somewhere that intel has AVX-512, which improves the performance of matrix multiplication which is important in ML.

[VanBantam]

I just built a deep learning rig using the following parts:

AMD threadripper 1950x 16c, 128gb non-ecc ram, 250gb nvme ssd for os and 1tb nvme for data staging, 2x2tb ssds for data storage, and 4 x gigabyte 1080 ti 11gb oc.  I'm running centos 7 no compatibility issues. 

I went threadripper because of the 64 pcie lanes.  I was looking at a dual xeon setup for a little while but couldn't find a reasonable priced mobo that had enough pcie slots with dual socket.  Plus also ecc ram is expensive.

 

The 1080 ti is the best cost per performance gpu right now.  

[daanav]

@VanBantam Thanks for replying. Yes the compatibility of TR has improved on Linux and I even read it is performing better on Linux as compared to Windows.

But I still have a question for you. Why did you go with Threadripper? Isn't i9 7900x faster than TR at Single core speeds and for deep learning, single core speeds are more important.  

[VanBantam]

6 hours ago, daanav said:

But I still have a question for you. Why did you go with Threadripper? 

I went Threadripper for PCIE lanes.  I want to run as many GPUs at x8 or x16 as my hardware can handle.  For my budget a Threadripper build made sense.

6 hours ago, daanav said:

Isn't i9 7900x faster than TR at Single core speeds and for deep learning, single core speeds are more important.  

I don't know about the performance specs for the i9 7900x compare to a 1950x.  With respects to cpu performance and deep learning my understanding is that a cpu does not perform as well as a gpu.  I'm sure there may be an edge case where cpus are preferred.  I do know that cpus outperform gpus for reinforcement learning.  However for let say computer vision a cpu just doesn't compare.

 

Search for Tim Dettmer's guide about choosing gpus for deep learning.