Re: Fwd: [Beowulf] Free Fermi cards to interested developers and researchers
Michael Clark <[email protected]>
| Newsgroups | gmane.user-groups.linux.singapore |
|---|---|
| Message-ID | <[email protected]> |
On Tuesday 27,April,2010 09:01 AM, Joel Wiramu Pauling wrote: > I would love to get hold of some for network > inspection/filtering/classification offloading. That would be interesting. As long as you can make everything fixed length and avoid branching (use predication) I think it could be done quite efficiently on the GPU. e.g. buffer up say 100K packets and run a kernel function over a whole batch, reducing the batch size and submission frequency based on network traffic. Deep packet inspection may be harder although not impossible (note this is pre-cuda but same concepts apply): http://http.developer.nvidia.com/GPUGems3/gpugems3_ch35.html Luckily Fermi can handle 32 bit integers (previous GT200 could only do 24bit integers which would make packet stuff harder - 32 bit integers are available in CUDA but are emulated and slow on previous gen nvidia GPUs). Anyone else on the list playing with CUDA? (i'd be keen to share experiences). _______________________________________________ LUGS Mailing list - [email protected] List FAQ: http://wiki.lugs.org.sg/LugsMailingListFaq Info page: http://www.lugs.org.sg/mailman/listinfo/slugnet To unsubscribe send an empty email to: [email protected]