OCZ RevoDrive (or similar) in peer-to-peer copy? Is this something that CUDA will support in future

Accelerated Computing CUDA CUDA Programming and Performance
1

Dear NVIDIA folk,

Recently at the SC11 conference I learned about PCIe-based SSD storage such as the OCZ RevoDrive and upcoming Z-Drive, that have transfer rates up to 2 GB/s.
My question is, since the card is already in a PCIe lane, is there some possibility CUDA can support direct memcpy to and from these devices as if they were a second Fermi GPU in peer-to-peer mode?

I imagine that it would be much faster than treating them as a standard SSD drive and having the data stream through the system bus?

Any thoughts/insight into this greatly appreciated.

Cheers,
Mike

2

Hello

I bought one revodrive 3 120 Go (not X2 series) and it rocks really espcecially with application using IO acces on disks.

The big CONS : NOT USABLE UNDER LINUX OS. there is no driver at all. One guy (not official) made a driver for it but OCZ said , they don’t plan to make any drivers for revodrive 3 External Image External Image External Image External Image External Image External Image

3

thanks for this feedback! that’s exactly why i want to use the Revo 3 X2’s External Image

I assume you can’t get CUDA to recognise it in p2p mode, even though is IS just another PCIe device??

what i’m wondering is whether (in p2p mode), Revo–>PCIe–>Tesla would be much faster than Revo–>PCIe–>RAM–>PCIe–>Tesla (as normal ssd storage mode).

have you done any raw IO benchmarks that you can share?

since if benchmarks show that you can saturate the Revo going the long way then who cares about p2p!

4

Yes sure. here it is

http://www.evga.com/forums/fb.ashx?m=1336334

That’s me Terminou on the EVGA forum

5

thanks. that looks fairly impressive even for a non-X2.

so there’s definitely no way to go direct across PCIe to global mem, right?
in CUDA, you still have to load into system ram and do a memcpy