On Wed, Dec 12, 2012 at 03:46:27PM -0500, Timothy Normand Miller wrote:
> It wouldn't be TOO hard to get some small FPGA or microcontroller, wire it
> to a set of flash chips, and hook that to a PCIe bus.  Then we'd want to
> use something like NILFS2 as the filesystem to handle wear-leveling in
> software.
> 
> We can just expose the raw flash organization to the host so that the host
> software can allocate blocks and schedule accesses to individual flash
> chips.  So we'd likely need a specialized filesystem that reorders requests
> so that blocks are pulled from as many chips as possible at once for
> maximum throughput.  It would also have to handle block remapping, garbage
> collection, etc.  We'd just put an accessible ROM on there that lays out
> the structure for the kernel driver.

How about this:

PCI-E form factor board, with:
* 1 CX4 Infiniband/10gigE connector, wired to FPGA
* x1, x2, x4 PCI-E gen1/gen2, wired to FPGA
* DIMM-type sockets for removable flash chip modules

It looks like the Xilinx XC6SLX75T can support 8 
GTP transceivers, that looks like it *might* support
both the Infiniband and x4 PCI-E at the same time.
(unless you need a 2 GTP's for a tx/rx pair)

Are there other FPGA's that would work for this?

(As for why infiniband, see http://seniord.ece.iastate.edu/may0904/ )

Let me add a crazy thought... Can we make GDDR5 'DIMMS' that could
be put in instead of flash chip dimms, and throw on a displayport 
connector on the same board? (I probably just added $1500 to the cost
for this craziness)
_______________________________________________
Open-graphics mailing list
[email protected]
http://lists.duskglow.com/mailman/listinfo/open-graphics
List service provided by Duskglow Consulting, LLC (www.duskglow.com)

Reply via email to