I expect to have my lab buy a Tesla C2075 some time since the FFT tests I managed to run on a C2050 were really exciting (512 cube single-precision in 0.12s). I simply can’t buy another Tesla board from my country due to university buying limitations: I can’t take some credit card and go to amazon, unfortunately.
The fact is: I need to perform real-time visualization of data computed from my Tesla/GTX. So I’ll have to make use of the DVI display on the back of my (future?) Tesla.
How does it work? Can it run simple OpenGL programs in background while the CUDA cores perform series of FFT?
Another question: if the C2050 and C2070 can handle 2560x2000 (?) display with ease, the C2075 can only handle obsolete 1600x1200. We have 1920x1200 monitors, and it seems sad to use a high-end GPGPU with either an old 1280x1024 LCD or a distorted image on a recent LCD.
One workaround would consist in plugging an additional 1280x1024 screen onto the Tesla and putting the 1920x1200 LCD on a cheap GTX card, but I can’t understand how Xorg (linux x86-64) will handle the display in this configuration.
In any case, there should not be PCI-express communications between the computing and displaying device: in this case, I’d probably better buy a GTX 580 with 3Gb…
Thanks in advance for any hint on one of my questions. Please feel free to drop remarks or just criticize my assumptions: I’m not a CUDA expert at all.
Yes, you can. In fact, for many workloads, it is better to have one GPU doing both the compute and the viz than to be juggling data transfers between multiple devices. For a great example of compute and viz mixing, check out our ArrayFire example demos (like the shallow-water equation example or the FDTD solver demos).
If I were you, I’d give it a try with the Tesla and see how much you can push it. Then decide if you need something else. Tesla’s are pretty capable for viz too.
I hoped some overworked intern at nvidia failed in typing the specs, actually. I don’t understand what they expect to spare on such a costless chip compared to the price of the whole setting.
To be honest, I might cancel the acquisition of the Tesla board if it turns out too complicated. I really hope not, but I’m not pushing so much money to get stuck afterwards.
I will eventually reproduce the current setting to a few acquisition PCs so I must keep it simple, with or without Tesla.
Tesla C2075 ships in TCC mode by default, which means that it operates in compute-only mode, resulting in up to 5-10% increase in compute perf depending on the application. In TCC mode, Tesla boards do not provide any graphics acceleration, meaning the resolution is limited 1600x1200 max, which is what MS Windows will allow.
User have the option to switch to WDDM mode using NVSMI software, which turns on hardware graphics acceleration and get up to 2560x1600 @60Hz default resolution. Tesla C2050 / C2070 shipped in WDDM mode by default.
I assume there are no big differences between Linux and Windows, since I expect to run my machine on linux as I did for testing the C2050 of my colleague.