Hardware Options for Quad / No CPU Memory and distributed storage by MCP

What hardware options for GPU processing exist that do not rely on an external CPU and related slow memory controllers?

The nForce MCP handles the pciE as well as “4 gbit ethernet” and has a htt link and sli, are there any physical products optimized for making effective use of the GPU boards (G80) that can function without a slow and annoying CPU?

Is there code and design for the MCP that makes use of the hypertransport links and communications without needing the additional mess?

There is reference to SLI memory controlled by the MCP, but no information that indicates the GPU can handle memory transfers either to ram or wire and moreso no technical specifications on throughput from the GPU to pciE and then outward.

Our typical design for other processor systems has been a simple hypertransport where a specialized communications controller handles access of mass distributed systems as well as a few rings of htt bridge/tunnel units to allow large shared memory banks and external comms transparent of the processor or memory system used.

Obviously putting an nForce controller on one of these links would make use of the comms AND sli-gpu capabilities, whereas there are many options for pciE controllers.

Is this a design function of the nForce and where are the technical specs and code for such?

Moreso, has anyone built nForce-only boards that handle the full Quad-GPU arrangement?

Im sure these questions will become more prevalent when consumers notice you can put 4 GPU boards in a common pciE bus and clip it to a fiber with faster throughput than the Tesla units and a fraction of the cost.

-Wilfred
WilfredGuerin@gmail.com

Cross reference to technical design comments:

http://www.gpgpu.org/forums/viewtopic.php?t=4598