Threadripper 2990wx workstation

[lacion]

Beautiful components. I wonder, though, how far sub 5Ghz it’s worth going on the primary driver system versus splitting the budget to a faster workstation and a server solution that scales with use.

my use case benefits a lot more with cores and thread than raw performance on single core (VM’s is not my primary use case.)

I’m kinda a newb, but what I ended up with is a 2700x main system (8 core 4.2ghz sustained boost no problem with a 280mm rad), a fileserver (another 8 cores with dual gigabit nic), and an old r610 (dual hexacore with the memory downclocked to reduce power consumption).

i already have a 2700x file server that runs a handfull containers for my inhouse use. 10GB nic on that one

I kinda wish my ryzen system had another 500MHz and 4 more hyperthreaded cores for non-cuda machine learning. When I run small batch tensorflow stuff (50 epoch on a 10kb file with say 10 parameters for cross val), it finishes pretty fast, but not quite fast enough for my taste. I also wonder if faster memory would be nice (i’m at 3.2Ghz, but that 4Ghz on Zen2 looks tasty).

my ML use case is very heavy my test datasets have millions of entries i need the gpu helping out.

when im done with tests and development long training seassing in prod or stage happen in the cloud

I kinda wish I had a 10gig internal network (just a small switch). When my 610’s proxmox hypervisor backs up the VM’s every day, it saturates a 1gig stream to my fileserver and I haven’t figured out the load balancing dual port thing yet. Me being a better system admin would greatly help that, but I almost would rather be dumb on a single 10gig connection and not worry about it so much.

the zenith extreme has a 10gb pci card, i have 2 other servers rocking a 10gb nic, for switching im using netgear with 4 rj45 10gb ports and 1 SF port Netgear XS505M

I REALLY wish I didn’t install CUDA on my workstation. Workstation graphics drivers suck donky for games. Not really bad or anything, but certainly not optimized. Frankly, I loaded my workstation with dual NVidia gpus to use tensorflow-gpu and months on I just don’t do it. If I need them, I rent a box online for a day after I proof the run on my CPU. Again, if I was a better config artist, less problem. But when you get those errors. And you aren’t sure where the bottlenecks are…I really wish I had wholly separate systems for some of the more complicated hardware configs. For CUDA. For Cuda. What am i talking about. I want to get a linux box with a bunch of cards in it, spend a week figuring out how to f’ing make it work, and not touch it except to run scripts!

i been doing for my production servers for a while, not an issue for me, also for gaming i plan to add a VM with passtrought one of the GPU’s

I wish I had more PCIe lanes. 20 isn’t enough. I get stuttering and weird balancing depending on how I plug in monitors, pass the GPUs around. Now I’m on a windows main for the workstation for PTC Creo stuffs, and that makes the problem even worse. Windows 10 Pro doesn’t handle lots of peripherals well and you can’t pass them to VM’s.

dont spect any problems here the TR 2990wx has 60 pci lanes

VNC Viewer and windows remote desktop have made it happen for me, though. I run critical software on windows VM’s (business side quickbooks for example) and make a remote desktop link on my workstation desktop. Load times are way faster to connect to an already open VM. With VNC Viewer, putty, and monitors, I can mostly monitor everything at once should I want to. However, I can’t pass data to them easily. I’ve mounted shared drives across the machines and can drag drop, but it’s clunky. I have to click 4 or 5 times and I think a single workstation I’d only have to click or CD/dir once. I’m also highly reliant on an always-on fileserver and application server, which isn’t super efficient…

i dont think thnis will be an easy for me, i will only have 1 vm with a dedicated GPU, and if i have more they will be most probably be server VM’s with no X or windows manager all managed to SSH

Just a thought. If it were me, I would get the 16 core and spin the savings off into a used 710 with a couple hdd’s in mirrored zfs and proxmox or vmware.

go for it, your use case is a lot less extreme than mine. i been using 2600 and 2700x for a lot of small home servers for me and my friends and they work pretty well specially with unraid.

[vbrandon]

Sounds like your spec is about as good as it gets. I saw the pic of the parts coming in. Super amazing. I am curious what kind of numbers it puts out. Plan to post any benchmarks?

If you ever want to start a thread on scheduling/loading GPUs in tensorflow/theano/pytorch, I’d be happy to read!