Cross-posting my reply from another thread since I think this table might be relevant to this discussion.
Recently, I was researching nVidia GPU cards and synthesized my leanings of the cards in a table here.
I always found nVidia’s naming super confusing (GTXs, Kepler, Tesla, Pascal, etc). So I made this table. I hope others find this summary in table format useful as well.
I’m currently in the final stages of building my own DL rig. I plan to have a 2080ti for now, and I’ll add a second video card in a few months. Any comment and suggestion on my build would be greatly appreciated !
I would go for a 1To SSD if you can afford to, especially if you intend to work on large datasets (ie. computer vision), plus you can use a small part of it to boost your RAM by allocating a part of it as a SWAP file under Ubuntu, like 64/128Go, way cheaper than RAM imho as it comes to Go per $.
Unless you have a strong reason to chose the Xeon, I’d go with either a Ryzen 7 8-core for less money, or Threadripper 8 or 12 core if budget is an issue now (and still get better perf vs. that Xeon) with easy update path to 16 or 32 core if you need to.
I agree, a late generation i7 or i9 is a better value, since muttiple threads are only useful in some problems. But a lot of RAM is useful so go for 64GB if you can.
I got around having 2x gpus with top fan intakes next to each other by using a pcie riser to move 1 out of motherboard. Was a fair bit of work though. Will move to watercooling though when get a third gpu, 3 x this style gpu (blower style too noisy for me) will be too hot for me even with all the fans i have and high airflow case.
All right so I’ve reworked it a bit. I originally choose a intel Xeon because I couldn’t find enough PCIe lanes on other intel CPUs (but in fact i7 - 6850K have 40 so I just didn’t look hard enough). In the end I settled for a AMD threadripper. Any comments on this ajusted build ?
Don’t pick that specific liquid cooler. For Threadripper (or more generally the TR4 socket) you want a cooler whose surface completely covers the chip, and a lot of liquid coolers supposedly “work” with Threadripper but just have an adapter so they fit but the surface area is not enough. With the 1900X you may not experience problems but if you upgrade that cooler will not be optimal. Even an air-cooler like Noctuas would equal if not surpass in performance that specific liquid one. Check this Noctua NH-U14S TR4-SP3 review. Re: mobo if your budget allows it, go for MSI MEG X399 Creation which is a X399 refresh, which would be better suited if you plan to move up to a higher Threadripper later.
Re: memory, pick RAM in 4 sticks (not 2). For Threadripper you want quad-channel.
Thanks for the tips !
May I ask what’s the difference between the two MSI mobos ? Is the upgrade really worth it considering most of the heavy lifting will be done on the GPU ?
The “refresh” mobo would only make sense if you plan to go with a higher-end 2nd gen threadripper later and if you then plan to overclock it. Upon further reflection: no ;-). Money would be much better spent in a 2nd GPU when the time comes.
Lambda deep learning workstation with 4 1080Ti uses Blower edition gpu. They say better for thermal management. The workstation also have fans in front and back of enclosure.
The 2080ti is slightly overpriced, you can find it online for less than 1100.
The threadripper is good, but consider that you get to give up upon intel MKL
I’d go for an Asrock x399 taichi, but this is based upon my personal experience with various mb brands…
Since you are getting a rig that allows for 3-4 gpus, maybe a 1200W psu could be a better choice. Conversely, if you plan to limit yourself to 2 gpus, something like 750/850w could allow you to enjoy better efficiency.
If you won’t go above 2 gpus, a 64-lanes rig could be overkill… You can save money about the cpu/mb and go straight with two 2080ti
I have 2 GTX 1080 ti Gygabyte on motherboard Gygabyte Z370XP SLI.
Linux UBUNTU 18.04.
I can use the 2 GTX boards indipendently, but if I enable the SLI with nvidia-xconfig --sli=On
at the successive reboot I obtain a black screen after the login.
Can you tell me exactly your configuration and how did you make work the 2 boards in SLI?
Thank you very much
I have 2 GTX1080ti Gigabyte and a motherboard Z370XP SLI Gigabyte.
I have some problem enabling the nvidia-xconfig --sli=On
Can you tell me which Linux version ans which driver you use?
Thank you very much
I don’t have SLI enabled. You don’t need it to do multi-gpu with fastai/pytorch. You just need to use nn.dataparallel. There are plenty of posts on how to do that. I haven’t been keeping up to see if in the last few months that there are advantages using SLI, but there didn’t use to be. I think NVLink (Similar to SLI) on some of the newer cards can take advantage of the additional direct card to card bandwidth it provides, but I have never used it as my cards don’t support it.
