Rather than using scatter/gather DMA, the GuPPI accepts the physical page addresses within large memory buffers allocated by a user-level device driver. This mode of operation is well suited to the transfer of a continuous sample stream. The GuPPI can transfer samples into host memory at rates up to 930 Mbps and from host memory at rates up to 830 Mbps.
A rough specification of the GuPPI can be found here.
Making Commodity PCs fit for Signal Processing, a paper about the GuPPI presented at USENIX'98.
For more information about the GuPPI, contact izzy@lcs.mit.edu