Jump to content
Search In
  • More options...
Find results that contain...
Find results in...
quantacide

Number crunching: 1st gen Threadripper vs. 9th gen Intel

Recommended Posts

Posted · Original PosterOP

I'm just starting to scratch the surface with some deep learning/Tensorflow machine learning and my current rig is... wait for it.. a 2012 Mac mini w/ a Samsung EVO SSD and 16GB of RAM. Yea, so I'm quite a ways behind.

 

After toying with whatever is on sale at Dell Outlet I really want to build my own rig again (last build was in 2008 with an Intel Skulltrail) and I'm enticed by the AMD Threadripper. However, my budget ($1000 +-) doesn't allow entertaining current v2 v3 Threadrippers. So the question is, do I build an "upgradable" rig or do I go team blue and just do an i5/i7 and then wait until I can get all those threads/channels. More threads means more crunching, and the "base" Threadripper allows me a lotta threads to expand (I don't think I can get to the same end where I have 4 GPUs w/ team blue) but who knows where/what will be out there once I CAN upgrade.  

 

TLDR: I want a good base to start building a real algo/deep learning/number crunching rig AND want upgrade flexibility AND as much bang for buck as possible using $1000. Do I go tail between legs and Intel i5/i7 or use borrowed 1970s F14 Tomcats and go AMD Threadripper v1? 

Link to post
Share on other sites

If you can get Threadripper 1950x + 4x4/4x8 GB of RAM + MoBo combo for around 600-700$, go this route. But I would actually recommend go Ryzen 3800x/3900x path + 2x8 GB of fast RAM (like 3200 or, even better, 3600 MHz). That way you can have more potential for upgrades than Threadripper 1st/2nd gen and Intel.

Plus get 1660 Ti or RTX 2060 SUPER. That can go long way for deep learning and TensorFlow.


Purify your Windows 10, don't give Microsoft anything that you don't want to share.

https://drive.google.com/open?id=1ZwVs9zrM493rjD42E2Pf0YcOkaW92ZUo

BTW, I am folding on laptop now, are you? https://stats.foldingathome.org/donor/Spakes

Tips for folding on laptop:

Lazy man wants upgrades from the sky.

Link to post
Share on other sites
Posted · Original PosterOP

I really liked starting down the Threadripper line b/c I can get so much more bandwidth from GPU/memory down the road (i.e. I have more lanes to add more cores) whereas I'm topped out on the Ryzen line. Is that the case? 

 

My upgrade concerns are more on the GPU side than on the processor side. I just want a good base. 

Link to post
Share on other sites

what the most important specs

cores?

ram?

cpu?

I have no clue what specs you want to optimize for a deep learning build. If you need a lot of ram. It might be worth it to look at X99 build because you find really really cheap DDR4 ECC ram on eBay and you don't need high speed ram because there is no infinity fabric speed to match. 

 

$1000 used build:

$175 E5-1660 v3 (i7-5960x, 8C/16T)

$200 X99 Motherboard 

$150 64GB DDR2133 ECC RAM (SAMSUNG M393A2G40DB0-CPB)

$450 GTX 1080 Ti

 

I know this setup would be much better for gaming compared to a threadripped v1 since the IPC is higher on haswell-e and it's cheaper too but it might not be as good for deep learning if you need more cores/threads. 

Link to post
Share on other sites
1 hour ago, quantacide said:

I really liked starting down the Threadripper line b/c I can get so much more bandwidth from GPU/memory down the road (i.e. I have more lanes to add more cores) whereas I'm topped out on the Ryzen line. Is that the case? 

 

My upgrade concerns are more on the GPU side than on the processor side. I just want a good base. 

The problem with 1st/2nd gen Threadripper is that old MoBos on X399 TR4 socket don't support newer lines off Threadripper that are built for TRX40 chipset and sTRX4 socket (same pin layout, vastly different pin config). Same goes vice versa. But if you want to get more GPU juice, than yeah, 1950X might worth a shot. Especially with those Titan Vs or Turing cards with tensor cores.


Purify your Windows 10, don't give Microsoft anything that you don't want to share.

https://drive.google.com/open?id=1ZwVs9zrM493rjD42E2Pf0YcOkaW92ZUo

BTW, I am folding on laptop now, are you? https://stats.foldingathome.org/donor/Spakes

Tips for folding on laptop:

Lazy man wants upgrades from the sky.

Link to post
Share on other sites
Posted · Original PosterOP

Good call on the Xeon/DDR ECC RAM. 

 

Number of GPUs is probably the most important. Hence why I've started at Threadripper (64 lanes). The alternative is to go a totally different route w/ fewer lanes and go Ryzen (16 lanes). 

 

I put together: 

 

$160 Threadripper 1900x

$280 Gigabyte X399

$140 32GB DDR4 3000

 

 

 

Link to post
Share on other sites
Posted · Original PosterOP

After doing more (and more research) I may be putting too much emphasis on the lanes. I think a Ryzen/Intel setup w/ a GPU or four would still work with the marginal benefit of having more lanes being small. 

Link to post
Share on other sites
Posted · Original PosterOP

Well I went and got a Ryzen 3900x. It doesn’t have enough lanes to bankrupt me with video cards but it’s more future proof than a 1st/2nd gen Threadripper. 
 

And now I’m team red!

Link to post
Share on other sites

I'm in a similar boat, with an R5 3600 and debating a 3950x or higher.

 

One thing that isn't mentioned enough with Ryzen is the Intel MKL library crippling on non Intel. There are workarounds, but python and other commonly used ML/DL frameworks utilize the MKL by default.

 

Performance is hugely crippled. 2-3 orders of magnitude in speed crippled vs the Intel chips, especially AVX. AVX-512 is even more of a difference.

 

Good news is, at the hobbyist level doesn't matter and can be worked around.

 

At the pro/getting paid level, it's one of the very few instances where a 10980xe or the other x299 10xxx chips are warranted-- out of the box avx (and, avx-512) support without any workarounds needed.

 

Do some google on the MKL with respect to AMD, employ the workaround, and enjoy!

 

Also, yea, GPU training is a thing but CPU running of trials before pushing to GPU is also important. As is running and evaluating the model after training.

Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now


×