The current memory requirements are going to exclude many people who only have one GPU and are looking to number crunch opportunistically with it while they are not using it for anything other than typical 2d display stuff.
In principle you are right. However, modern desktop environments almost always use 3D acceleration for compositing (overlays, shadows) and window animation. Thus the only reasonable way to use the primary (display) GPU for crunching is while you're not using the machine interactively - otherwise you'll experience the aforementioned sluggish response.
Regarding the memory requirements: this has been our very first CUDA release, hence we didn't make any significant changes to the search algorithm. Future versions might be designed differently such that they process data in a different way, requiring less GPU memory. It remains to be seen, however, how these changes are going to be prioritized (if we consider them worth the effort).
Quote:
Well, it's a fairly vanilla Imac. The only thing the card is doing is providing display on the primary Imac monitor. I'm not using any GL or GPU applications other than what MacOS uses for display acceleration (not even a 3d screensaver).
I'll have to do a reboot to see if the problem persists and how long it persists.
See my statement above. An additional problem might be that your NVIDIA device doesn't have dedicated GPU memory but uses parts of the OS RAM under a best-effort policy. The same issue can be observed in Laptops using on-board GPU chipsets. Regarding the current 21.5" iMac models for example: only the ATI card sports dedicated video memory while the NVIDIA chip doesn't.
Quote:
The Cuda app started after a reboot. I have not a clue what could be taking up GPU memory inbetween the time a cuda app ends and a new one begins. Should it release all the memory it was using before proceeding onto the next app?
Well, that's true for GPU (i.e. CUDA) applications as such but not necessarily for the OS using 3D features as mentioned above. The amount of memory occupied really depends on your actual usage, e.g. the screen resolution, enabled eye candy features, monitor setup, etc. or maybe even some image/texture remnants - and all of this gets worse when your GPU doesn't have dedicated memory...
RE: The current memory
)
In principle you are right. However, modern desktop environments almost always use 3D acceleration for compositing (overlays, shadows) and window animation. Thus the only reasonable way to use the primary (display) GPU for crunching is while you're not using the machine interactively - otherwise you'll experience the aforementioned sluggish response.
Regarding the memory requirements: this has been our very first CUDA release, hence we didn't make any significant changes to the search algorithm. Future versions might be designed differently such that they process data in a different way, requiring less GPU memory. It remains to be seen, however, how these changes are going to be prioritized (if we consider them worth the effort).
See my statement above. An additional problem might be that your NVIDIA device doesn't have dedicated GPU memory but uses parts of the OS RAM under a best-effort policy. The same issue can be observed in Laptops using on-board GPU chipsets. Regarding the current 21.5" iMac models for example: only the ATI card sports dedicated video memory while the NVIDIA chip doesn't.
Well, that's true for GPU (i.e. CUDA) applications as such but not necessarily for the OS using 3D features as mentioned above. The amount of memory occupied really depends on your actual usage, e.g. the screen resolution, enabled eye candy features, monitor setup, etc. or maybe even some image/texture remnants - and all of this gets worse when your GPU doesn't have dedicated memory...
Cheers,
Oliver
Einstein@Home Project