All current platforms can do DMA between various IO devices and AGP directly to main memory, while cpu is doing its cpu stuff. I don't know how exactly DMA works on Hammer. I suspect that technically everything is ok, because, after all, Hammer is designed to cross-switch memory requests from remote cpu's to local memory. But I suspect this process is so much different than DMA, that developers must really scratch their heads. And, I know it's absolutely impossible to make productive hardware and software development until you have 3-5 real computers in the lab. Most likely Hammer chipsets are pretty crappy at this moment.
Market Data 
Markets 
AI
