CUDA 5 can share memory?

[Dariusz Makowski (Dadal)]

Heya

I've ben just briefly reading about cuda 5 and it looks like each GPU can access other GPU memory, is it true?

Any comments from chaos?

Ty, bye.

[super gnu]

wow that would be a game changer. 4x 4gb gpus trouble is, nvidia give with one hand and take awasy with the other.. since the kepler cards are crippled deliberately. wont be buying 4 quadros or 4 teslas thats for sure.

[Dariusz Makowski (Dadal)]

Not exactly for precise answer u need some1 with engineering skill but because Keplar has changed quite a lot it could be a software issue that render enginers need to adress. The new SMX, hyper Q and dynamic parallelism might be what has to be adressed in vray in order to get max out of Keplar GTX... not sure need chaos answer

Thanks, ciao.

[vlado]

I've ben just briefly reading about cuda 5 and it looks like each GPU can access other GPU memory, is it true?

I have not looked in detail into this yes, but my guess is that you won't get one large uniform address space; instead it will still be necessary for programs to explicitly copy blocks of memory from one device to another...

Best regards,
Vlado