Will Geforce 8200 (integrated) support CUDA?

I’ve seen some material (at Anandtech) about G8x-level integrated graphics. I’m curious if this will mean that there will be a CUDA-capable card that has direct access to some portion of main memory. Even if this isn’t a beefy card, it’s still an interesting possibility, as in many cases, it’s our transfers to and from the GPU’s memory that are our main source of latency, not the processing.

I could see this going either way, though. Perhaps making CUDA work on a IGP is too much effort for too little reward. I know that NVIDIA has said that they’ll make CUDA work on all 8xxx-series cards and beyond, but the statement wasn’t really clear about whether it would apply to integrated graphics.

Thoughts?

We haven’t announced anything about integrated graphics support, but as you say it is our plan for CUDA to be supported on all G8x parts.

It is actually possible for G8x GPUs to texture from host memory (and in fact the hardware has been able to do this since the AGP days), but the performance is relatively poor.

I’d be curious to hear how useful this would be to people, and what kind of performance would be tolerable.

For some tasks, the having the equivalent of a low-end cards (current 8400-8600) in IGP and not requiring expensive transfers would give us a factor of 2-3x speedup, as we spend far more of our processing on transferss and take a big hit from startup latency.

The poorer performance texturing from host memory would probably be won back by not having to do the transfers.

Why were the calls that allowed CUDA to access main memory removed, anyhow? I know they were slow, but there were some extremely useful things that we could have done with them that would have allowed us to get around transfer and startup latencies…

It’d certainly be nice if CUDA could texture from host memory, or even read and write it directly if it’s page-locked. I have the suspicion this is already possible, just not exposed through the API.