[Hardware] partially unrolled?

John L. Bass jbass at dmsd.com
Mon Nov 27 19:23:45 EST 2006


ABC> I read on this list that a fully unrolled design is too large for any
ABC> FPGAs we can afford, while a fully iterated is probably too slow to be
ABC> useful.

Hi Guys,

I did some poking around at Xilinx FPGA pricing for devices big enough to
handle a fully unrolled pipelined design, and found that Spartan-3 devices
are pretty affordable and have enough features to realistically support a
standard Dnet design.

XC3S2000 devices are just under $100/each, which means it would be practical
to design a very low end PCI card that supported up to 8 of these devices.

With some pruning, I suspect that we might also be able to fit the core into
an XC3S1500 device which are price about $75/each in hundreds.

The XC3S1500 device has the added benefit that it's supported by the free
web pack tools, meaning that hobbiests would not have to invest in the
$2K Xilinx ISE tools to use their accelerator board.

I suspect the performance will be near that of the Virtex-4 devices, and
much better on a MegaKey/Sec per Dollar scale.

Guerric should have better numbers in a day or two, and we are collaborating
on improvements and strategies, off list.

Both of these devices would make practical general purpose FPGA computing
engines using available C compilers, as well as traditional VHDL/Verilog
designs. Using low overhead production, boards priced between $150 for a
single FPGA and about a grand fully loaded are very possible. If done
commercially, they would have to carry a 2x or 3x markup to cover other
costs.

The parts are only available in fine pitched BGA, which will increase board
costs slightly due to the small vias required and extra tight layer to layer
layup error management. There is a lead-free program for the parts, so we
can be environmentally friendly too.

----------

Can someone post what the Dnet core counting sequence, work assignments,
and work checking requirements are for a valid DNet core to participate?

I'd like to work with Guerric to get a functional first pass core running
on silicon just to validate the design and do some power/thermal testing.

Have Fun,
John


More information about the Hardware mailing list