[tech] [Conversation Starter]: GPU server

[tec]

Hi All,

With the general rise in GPU-accelerated compute tasks, and particularly ML I think it could be a good idea to be able to have VMs with access to GPU resources.

Give me your thoughts!

Also, here's some relevant copy-pasta from discord:______________________________________________________________________________
tec28/09/2019I'm looking to try to install a bunch of stuff to try out GPT2, and hopefully not break anything. Have we got any GPUs I can access via a VM? I would install on my desktop, but my distro doesn't have the nvidia cuda stuff packaged ______________________________________________________________________________
coxy28/09/2019loveday has some shitty nvidia card bah, can't get the dvb tuner working on windows10 ______________________________________________________________________________
tec28/09/2019 Just wanna check gpu ______________________________________________________________________________
coxy28/09/2019think you need to ssh in as root, check uccpass when we get the new proxmox hosts we should put something in heathred ______________________________________________________________________________
NickBOT28/09/2019 at tec @DylanH[333]👌: If you're up for it, that would be pretty handy and I'm sure there would be club funding available for that sort of thing (and/or donations?). "HowToGPGPU at UCC" ______________________________________________________________________________
tec28/09/2019I don't know how yet, but I'd like to ______________________________________________________________________________
DylanH[333]👌28/09/2019I'd attend a session on something like that, but I' m not knowledgeable enough to run one I could potentially help with GPU passthrough in ProxMox though I've done PCIe passthrough before for a NIC ______________________________________________________________________________
NickBOT28/09/2019 at DylanH[333]👌: Yep, step one is expanding the cluster and hardware installation - happy to help experiment ______________________________________________________________________________
DylanH[333]👌28/09/2019Yeah, I need to get back onto that. I still haven't taken the time to ask committee about ordering SSDs I'm now of the mind: didge the SD cards and reinstall Proxmox on 500GB+ SSDs, and dedicate an LVM volume for Ceph I just haven't had the time/motivation ______________________________________________________________________________
NickBOT28/09/2019As for the OpenCL? software setup, @tec will no doubt help? ::-) ______________________________________________________________________________
DylanH[333]👌28/09/2019Also, we'd have to use one of our existing hosts in the cluster for a GPU, unless we have one lying around that's no thicker than one slot ______________________________________________________________________________
NickBOT28/09/2019Sounds good to me, one big drive per machine now that we have >3 hosts? though the SD boot is neat, too. ______________________________________________________________________________
DylanH[333]👌28/09/2019The 1RU HPs have a full-sized PCIe slot, but it's pretty thin ______________________________________________________________________________
tec28/09/2019PCIe extension ribbon? ______________________________________________________________________________
DylanH[333]👌28/09/2019Still doesn't solve where to actually fit the GPU ______________________________________________________________________________
NickBOT28/09/2019Loveday should be good for physical space. ______________________________________________________________________________
DylanH[333]👌28/09/2019 at tec This is the sort of clearance the HPs have to work with ______________________________________________________________________________
TRS-8028/09/2019also how do you get power to the gpu? ______________________________________________________________________________
DylanH[333]👌28/09/2019PCIe slot - so it'd have to be 75W or less iirc I suppose most GPUs that fit that power limit will probably fit that size limit as well 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: https://lists.ucc.gu.uwa.edu.au/pipermail/tech/attachments/20191024/7c5a9042/attachment-0001.htm