I’ve got an HP Proliant DL380 Gen9 server with two slots available.
I’ve got around $15K Canadian available to purchase two cards for the system.
I’ll be working with Ubuntu16.10, Python, Tensorflow, Pandas, Theano, Sci-kit learn.
I’m doing some R&D on neural nets for a business applications so I’d like to get a Tesla for the first available slot. This chews up around ~11K so I have $4K left for a graphics card.
I’m also doing some graphing with up to millions of points in matplotlib -lots of points in 3d, so a high end card won’t be wasted…
I’m thinking a Quadro, but which one rounds me up to $15K?
Can anyone see a reason why running a Tesla P100 and a higher end Quadro on the my server, with the named software is going to be a problem? I’d hate to end up with a gotcha somewhere that I’ll regret.
Unless memory fails, NVIDIA themselves are pushing Tesla/Quadro combos, so I would not expect any trouble. It has been years since I have run Tesla + Quadro in the same system myself.
Since HP will integrate the system (*), it would probably be best to check what they offer in terms of Quadro GPUs for the remaining $4K. Be prepared to find a sizeable markup compared to what you would pay on the open market.
(*) Actually, after re-reading your question, it isn’t clear who will do the integration. Integration of high-end Teslas into machines by end users often leads to problems, as one can see from numerous questions about such scenarios (in particular involving the K80) in these forums. Best I know, from NVIDIA’s perspective such home-brew systems are unsupported.
Understood, as I’ve stated, I have the server already and would be installing the cards myself.
The NVIDIA website shows that HP will integrate an M60 into the Dl380 server from the factory - its listed as a “qualified server configuration” on the NVIDIA site.
The P100 is not listed as a “qualified” configuration with my server.
My thinking was that this was likely as much marketing as anything else, am I being foolish?
More than marketing, I would say. High-end Teslas often have specific technical requirements (e.g. memory apertures provided by the system BIOS) which cause problems when people try to integrate themselves. The fact that the P100 is not listed as qualified for your server would appear to at least hint at such issues.
I would choose cards that are qualified in the HP DL 380 G9. You should be able to get the list of available/qualified/supported cards from HPE. With that list in hand, finding out which Quadro card will “round the price up to $15K” should be an easy matter, if not I’m sure HPE sales can help.
You’ll want to make sure you review compatibility with HPE. Many servers have things like GPU riser kits/install kits which are not shipped with the server if it didn’t come from the factory with GPUs installed. If you intend to acquire GPUs on the open market and install them yourself, its possible you may have some missing pieces.