Unified memory on Cuda 8

[Sushidelic]

Any plans going towards this direction?
Unified memory / GPU memory oversubscription sounds awesome to me...

http://on-demand.gputechconf.com/gtc...ark-harris.pdf

Best regards, Michael

[savage309]

Any plans going towards this direction?
Unified memory / GPU memory oversubscription sounds awesome to me...

http://on-demand.gputechconf.com/gtc...ark-harris.pdf

Best regards, Michael

Sure. However, those will concern the GP100 Pascal core (Tesla P100). GP102, GP104 and GP106 (1080, 1070, 1060, etc) are quite different type of Pascal (which are very close to Maxwell) and those won't work for them. I think the GP100 could be a little pricy, at least in the beginning.

Best,
Blago.

[Sushidelic]

Ahhh, understood... by quickly reading it, i thought it concerns all kind of pascal chips, including 1080 or the new Titan X. Then actually there's not that much interest from my side, as bang for the buck unfortunately is still an issue, until we win the state's lottery

[savage309]

Ahhh, understood... by quickly reading it, i thought it concerns all kind of pascal chips, including 1080 or the new Titan X. Then actually there's not that much interest from my side, as bang for the buck unfortunately is still an issue, until we win the state's lottery

I know that feel. At the time of releasing these nobody knew that there would be two types of Pascal architectures .

Best,
Blago.

[padre.ayuso]

I think I understand this thread, but in layman terms??? That means that Unified Memory (which I'm a bit blank after looking at that link) is not possible except in some more costly cards?

I also overheard somewhere that the Pascal cards allow for memory stacking, which means that if you have two 8GB cards (1080's) you will essentially have 16GB? I know Blago told me not yet, but I thought this was because VRay would support it in the near future? Please help me clarify, I feel a bit lost on this one.

Thanks,
Alex

[savage309]

not possible except in some more costly cards?

Yes (well, we will know if it will actually work for us once those GPUs are shipped and we can test them).

I thought this was because VRay would support it in the near future.

No, it is not because of V-Ray not supporting it (yet). According to the white papers the GP100 (Tesla P100) will be the core supporting that, GP102 (Titan X Pascal), GP104 (1080, 1070), GP106 (1060), etc will not (GP100 is quite different and should be even better than those). We still have to add support for that for P100 and we will most likely do that when it is out.

Best,
Blago.