Network card

See http://daniel.haxx.se/blog/2008/12/18/10g-and-direct-cach...
It's also known as I/OAT.

To quote bgoglin:

> DCA doesn’t really reduce the memory bandwidth requirements since the
> data still has to be fetched by the cache from the main memory (the
> device doesn’t write into the cache, it just tells the cache that data
> should be fetched). The whole point of the approach is that this fetch
> is done in advance, so you don’t have to wait for it when the host starts
> processing the packet.

So it doesn't seem that good yet.

With integrated memory controllers I would think it's easy to do this automatically, or at least very easily, by directly going to the cache first instead of through RAM (at least the L3 cache).