I made a test environment with the following hardware/software details:
DL 580 G7 (I know it is not certified Server for nvidia grid, I only want to test the performance)
2 x E7 4870 2.40 GHz 10 cores
Nvidia GRID k1
2TB 10k HDD local storage
Nvidia latest drivers
My goal is to deliver good BricsCAD performance to the end users.
After successful deployment of vGPU in this test server, I create my first manually pool with vGPU enabled.
All seems to work ok from drivers perspective (nvidia-smi ok,driver loaded succesfully,GPU-Z,DirectX diagnostic tool ok).
I see better user experience overall in the VDI, but BricsCAD seems not to work well, for example during copy paste of an especially large object, the object is disappearing until paste is done and mouse cursor is lagging sometimes(already used absolute and relative mouse option)
I made some tweaks regarding BricsCad options (use hardware acceleration) and nvidia control panel but nothing so far give better performance.
Some points that can be affect performance:
1)DL 580 g7 is not certified (I was surprised that vGPU worked), can this be a problem regarding performance outcome?
2)At the moment, Grid k1 placed in PCI x16 slot, BUT this slot operates in x8. Maybe I lose some performance from that?
3)Grid k1 vs k2 have major performance difference regarding cuda cores (768 vs 3,072). Can cuda cores affect BricsCad operation?
4)low CPU frequency?
I already used all profiles but I don’t think that more memory in k180q can change something in performance.
I have tested OS Windows 7,8.1,10,Server 2012 with 6GB RAM and 4 or 6 cores
Thanks and I appreciate any advice,
Anyone have suggestions?
I made another test with vDGA enabled instead of vGPU and I notice some better performance on BricsCAD regarding mouse cursor. Has vDGA enabled features like CUDA where vGPU does not have?
Thanks and I appreciate any advice
Yes on GRID 1.0 K1 / K2 generation cards CUDA and OpenCL were not available via vGPU and only pass-through.
In GRID 2.0 and up M60 / M6 / M10 cards CUDA and OpenCL are available but only for vGPU profiles which are equal to the physical GPU i.e. Mx8-Q profiles.
In general for CAD the K2 (and now M60) were the recommended choice. The K1 was targetted at boosting VDI experience for Windows Aero / office apps / browsers. The K2 spec was essentially 2xK5000 GPUs, many serious CAD users have a dedicated K5000 to themselves, so the K1 (similar spec to K600) is below the spec many CAD users have in workstation even before shared.
The CPU can make a HUGE difference as many CAD operations still very CPU hungry. The copy and paste behaviour - you might want to check if the framebuffer of the vGPU profile used is well sized relative the size of parts you are using (avoiding paging like effects).
Thank you very much for the useful information.
So, more Cuda cores in K2 have performance impact in vGPU on CAD usage.
The only aspect I want to look for is that I used an non-certified Server. Can this affect the performance
or a certified server is about optimal cooling and properly seating these cards?
Thank you in advance.
With uncertified hardware all bets are off… it’s mainly about thermal/electrical stability but also about checking no PLX faults etc… it could work/it might not whether there are performance gremlins… no idea…
You really need to talk to HP/VMware and ask why they chose not to certify that server… and they can be open to certifying if there is sufficient customer demand.