Machine learning workstation & gaming.

[sybrscr1ber]

Hi all,

I might sound like I don’t how to google stuff like this but that’s the case here.

I spent hours doing my research and still can’t figure out the best setup.

Threadripper 1950x or i9 7900x? Ryzen 7 1700x\1800x or i7 8700k?

Motherboards is just whole nother story. Full linux compatibility / decent iommu groups.

I’m lost guys.

I’m looking for a good “under $4000” Machine learning / passthrough / gaming setup (gaming is least important).

Thanks

[Radium_Angel]

You think speccing hardware is a pain, have fun setting up the software for a RNN. Installing CUDA alone is a monumental PITA. You comfortable in the terminal? You're gonna need it, no fancy GUI for RNNs, it's all CLI

and if you are going RNN, intel/nVidia is much more mature than AMD/ATI (I am speaking from painful experience as a RNN junky)

[Guest]

I would get the threadripper 1950x and the Nvidia Titan V

[LtStaffel]

1 minute ago, Some Random Member said:

I would get the threadripper 1950x and the Nvidia Titan V

Why Titan V?

[Guest]

1 minute ago, LtStaffel said:

Why Titan V?

Nvidia touts it as the new machine learning champion, and if he is a baller he might just get it, also it has around 20% better performance than a 1080ti in gaming. 

[Crunchy Dragon]

1 minute ago, LtStaffel said:

Why Titan V?

It was designed for deep/machine learning, but at this budget it just wouldn't be plausible or worth it to dump that much of the budget into the GPU.

[sybrscr1ber]

6 minutes ago, Radium_Angel said:

You think speccing hardware is a pain, have fun setting up the software for a RNN. Installing CUDA alone is a monumental PITA. You comfortable in the terminal? You're gonna need it, no fancy GUI for RNNs, it's all CLI

and if you are going RNN, intel/nVidia is much more mature than AMD/ATI (I am speaking from painful experience as a RNN junky)

After being a full stack web developer \ devops since 2012 I decided to focus on ML \ upgrade my current hardware. So i'm ready for all those abbreviations you mentioned, lol)

[Streetguru]

20 minutes ago, sybrscr1ber said:

Hi all,

I might sound like I don’t how to google stuff like this but that’s the case here.

I spent hours doing my research and still can’t figure out the best setup.

Threadripper 1950x or i9 7900x? Ryzen 7 1700x\1800x or i7 8700k?

Motherboards is just whole nother story. Full linux compatibility / decent iommu groups.

I’m lost guys.

I’m looking for a good “under $4000” Machine learning / passthrough / gaming setup (gaming is least important).

Thanks

Probably want a 1950X for the PCI-e lanes

Then a Vega FE if you're going for freedom, or a Titan XP I guess, they're still available at MSRP.

[sybrscr1ber]

3 minutes ago, Streetguru said:

Then a Vega FE if you're going for freedom

What do you mean by "for freedom"?

[Radium_Angel]

12 minutes ago, sybrscr1ber said:

After being a full stack web developer \ devops since 2012 I decided to focus on ML \ upgrade my current hardware. So i'm ready for all those abbreviations you mentioned, lol)

Excellent.

Your hardware choices are then made for you. Intel CPU and nVidia GPU.

Personally I use Ubuntu 14.04 LTS because it still has the real nVidia driver installed and setting up CUDA and all the other nightmarish requirements are much easier because of that.

[Streetguru]

10 minutes ago, sybrscr1ber said:

What do you mean by "for freedom"?

Well if you buy into Nvidia you'll be stuck with nvidia until the end times and buying into a proprietary software ecosystem if I'm not mistaken.


Where as AMD generally supports open source stuff
https://instinct.radeon.com/en/the-potential-disruptiveness-of-amds-open-source-deep-learning-strategy/
 

https://github.com/openai

[Radium_Angel]

Street: Except RNN work is heavily based on nVidia, so you get to set aside your POV if you want to do Neural Networks.

I spent a great deal of time trying initially to get a RNN set up on ATI hardware, and it's a LOT easier on Intel/nVidia, the software platforms are far more mature, there is actual support for it, and so forth. It's really no question. You want to do NN programming, you go intel/nVidia.

[sybrscr1ber]

2 minutes ago, Streetguru said:

Well if you buy into Nvidia you'll be stuck with nvidia until the end times and buying into a proprietary software ecosystem if I'm not mistaken.


Where are AMD generally supports open source stuff
https://instinct.radeon.com/en/the-potential-disruptiveness-of-amds-open-source-deep-learning-strategy/
 

https://github.com/openai

I'm definitely an open source guy. I'll look into that. Thanks)

[sybrscr1ber]

3 minutes ago, Radium_Angel said:

Street: Except RNN work is heavily based on nVidia, so you get to set aside your POV if you want to do Neural Networks.

I spent a great deal of time trying initially to get a RNN set up on ATI hardware, and it's a LOT easier on Intel/nVidia, the software platforms are far more mature, there is actual support for it, and so forth. It's really no question. You want to do NN programming, you go intel/nVidia.

Yeah, makes sense. Have any idea what motherboard to choose? If Intel then should I go with x299 platform?

[Streetguru]

4 minutes ago, Radium_Angel said:

Street: Except RNN work is heavily based on nVidia, so you get to set aside your POV if you want to do Neural Networks.

I spent a great deal of time trying initially to get a RNN set up on ATI hardware, and it's a LOT easier on Intel/nVidia, the software platforms are far more mature, there is actual support for it, and so forth. It's really no question. You want to do NN programming, you go intel/nVidia.

I'm sure it is, if only AMD were smart enough to actually fight back in software support for AI stuff. But aside from that we'll be stuck with nvidia forever if people keep supporting them of course.

 

 

3 minutes ago, sybrscr1ber said:

I'm definitely an open source guy. I'll look into that. Thanks)

I edited the post, Level 1 Linux and Level 1 Techs has been working with threadripper a lot
 

 

[Radium_Angel]

Just now, sybrscr1ber said:

Yeah, makes sense. Have any idea what motherboard to choose? If Intel then should I go with x299 platform?

I use a core i5 on a stock Asus build, the GPU is the key for neural networks. Remember, linux isn't going to support SOTA hardware as well as Windows will, so beware choosing cutting edge hardware, esp. if you go with 14.04 LTS...

 

[Radium_Angel]

2 minutes ago, Streetguru said:

I'm sure it is, if only AMD were smart enough to actually fight back in software support for AI stuff. But aside from that we'll be stuck with nvidia forever if people keep supporting them of course.

Right tool for the right job. If AMD/ATI isn't interested, that's on them, not you. I have a RNN running under AMD/ATI hardware at work, it's *much* slower than my nVidia rig, and was a LOT harder to set up because the tools offered by AMD/ATI aren't there.

 

The other showstopper here is the OP wants to do gaming. While Steam does exist for Linux, only a small fraction of games run under Linux, and many of the ones on Steam are "visual novels" or very bad ports of 15 year old games.

 

You want to game, you might be thinking dual booting, or build a separate gaming box...

[sybrscr1ber]

3 minutes ago, Radium_Angel said:

Right tool for the right job. If AMD/ATI isn't interested, that's on them, not you. I have a RNN running under AMD/ATI hardware at work, it's *much* slower than my nVidia rig, and was a LOT harder to set up because the tools offered by AMD/ATI aren't there.

 

The other showstopper here is the OP wants to do gaming. While Steam does exist for Linux, only a small fraction of games run under Linux, and many of the ones on Steam are "visual novels" or very bad ports of 15 year old games.

 

You want to game, you might be thinking dual booting, or build a separate gaming box...

It's more about ML then everything else.

Is cutting edge hardware gonna be a problem? Major hardware announcements were made almost a year ago.

[Radium_Angel]

6 minutes ago, sybrscr1ber said:

It's more about ML then everything else.

Is cutting edge hardware gonna be a problem? Major hardware announcements were made almost a year ago.

With the latest and greatest Ubuntu, probably not. But it presents it's own share of issues with ML (depending on the style you want to get into)

All the RNNs I've worked with and set up and read about, use 14.04 LTS, which doesn't always support cutting edge hardware fully. 

[Streetguru]

47 minutes ago, Radium_Angel said:

 

 

42 minutes ago, sybrscr1ber said:

If I'm not mistaken the latest tech "Looking Glsss" Doesn't even require a 2nd GPU to do GPU passthrough, for gaming on linux
 

 

[brob]

59 minutes ago, Streetguru said:

Probably want a 1950X for the PCI-e lanes

Then a Vega FE if you're going for freedom, or a Titan XP I guess, they're still available at MSRP.

I don't understand how the distribution of PCIe lanes makes a difference. Both Intel and AMD platforms discussed here have about the same number of PCIe 3.0 lanes.

 

If a gpu is used for most of the deep learning processing, is a high core count cpu required as well? Mostly wondering if a Titan V with an i7-8700 or Ryzen 7 1700 would offer decent performance.

[Streetguru]

15 minutes ago, brob said:

I don't understand how the distribution of PCIe lanes makes a difference. Both Intel and AMD platforms discussed here have about the same number of PCIe 3.0 lanes.

 

If a gpu is used for most of the deep learning processing, is a high core count cpu required as well? Mostly wondering if a Titan V with an i7-8700 or Ryzen 7 1700 would offer decent performance.

Isn't it 44 PCI-e lanes for X299 and 64 for threadripper? Just means you get more GPUs at full speed no? Or more NVME?

Those are lanes directly to the CPU aren't they? not counting chipset?
 

 

[brob]

@Streetguru

 

Both platforms have roughly the same number of PCIe 3.0 lanes. The difference is in where they are located and how they can be used. It's true that Threadripper can support three full speed (x16) gpu slots vs two supported by i9. If one were contemplating a 3/4 gpu system, Threadripper would likely be the better choice. In this case it doesn't really matter.

[sybrscr1ber]

4 hours ago, Radium_Angel said:

I use a core i5 on a stock Asus build

Can you tell me what your exactly your hardware configuration is? I'm just wondering.

 

Also, what Nvidia GPU should I look into?

[Streetguru]

1 hour ago, sybrscr1ber said:

Can you tell me what your exactly your hardware configuration is? I'm just wondering.

 

Also, what Nvidia GPU should I look into?

The Titan XP is the only one in stock, and VRAM may be important, even that 1 GB over a 1080ti which is the same price right now anyways

 

Or the titan V if you can fit it into the budget, might be doable with an R7 1700, depends on how much you value your CPU.