Fermi and direct DMA from another PCIe Cards

[moozoo]

In the FAQ it states that it is not currently possible to DMA directly into GPU memory from another PCI-E device.

I read that Fermi supports this for particular InfiniBand drivers.

Is this possible for other drivers. For example high speed data acquisition cards?

I was looking at the QS1R VERB digital down conversion (DDC) receiver (http://www.philcovington.com/QuickSilver/) and thinking that a USB 3.0 version (if it was ever made) could dump the raw 120MBs data stream directly to the GPU which would then implement a full software based (MegaChannel?) receiver.

It should be able to do full 55Mhz wide spectrum displays and who knows what else (Super CW Skimmer? Identify and catalog every received signal?).
ie. As for http://www.youtube.com/watch?v=nae_M_q-i10 but 55Mhz wide and with 7hz(?) resolution.

[avidday]

In the FAQ it states that it is not currently possible to DMA directly into GPU memory from another PCI-E device.

That is correct

I read that Fermi supports this for particular InfiniBand drivers.

That is what the specifications suggest. The product is probably 4-6 months away from release, so nobody yet knows for sure (or if they do, they are either under NDA or work for NVIDIA and the information is not public).

Is this possible for other drivers. For example high speed data acquisition cards?

See above.

I was looking at the QS1R VERB digital down conversion (DDC) receiver (http://www.philcovington.com/QuickSilver/) and thinking that a USB 3.0 version (if it was ever made) could dump the raw 120MBs data stream directly to the GPU which would then implement a full software based (MegaChannel?) receiver.

I very much doubt a standard USB 3.0 controller (certainly any of the current crop of discrete controllers currently entering the market) would be able to do DMA transfer into the memory of another PCI-e bus device like a GPGPU card. In most implementations, those hang off the “low speed” I/O hub and don’t sit directly on the PCI-e bus. Further, the USB 3.0 protocol doesn’t have any of the memory mapping over the wire type features that native Infiniband verbs does (which is how I assume the DMA with Fermi will work - it is certainly how the Hypertransport bus based DMA adaptors for Infiniband and other exotic low latency links work).