Slashdot Mirror

DRC is one. http://www.drccomputer.com/index.html

if you are memory bandwidth limited or do too much of interleaved parallel & serial processing then you might find these folks ( http://www.drccomputer.com/ ) interesting. they do a opteron socket based co-processor. In many applications i have seen being able to model the right type of operation while hiding access latency works out best. Now only if somebody will does the same co-processor with nvidia gpus that would be superb.

you might be getting ripped off if you're paying $1500 for a Spartan-3 board.

I guess they don't really have the board volume to get low prices. But If you want a graphics card for $1500 that's probably less functional than an NVidia commodity card, I'm not gonna stop you.

OTOH, If you're interested in FPGA programming and a novice at it, you'll want to get a MUCH MUCH MUCH cheaper Spartan board (like 50 to 150). See http://digilentinc.com/ for good starter boards.
If you're serious about FPGA programming (or just willing to pay $1500 to $3000) you will definitely want to get a board with a Virtex or Stratix on board:
http://www.xilinx.com/products/devkits/HW-V5-ML501-UNI-G.htm

If you want to have it on PCIx:
http://www.xilinx.com/products/devkits/HW-V5-ML555-G.htm

You can also get FPGAs socketted for AMD's Hypertransport bus and Intel's FSB:
http://xtremedatainc.com/ (Altera FPGAs)
http://drccomputer.com/ (Xilinx FPGAs)
http://nallatech.com/
http://celoxica.com/

(some of these vendors also sell PCI solutions)

FPGA programming environments still mostly suck. it's a market impeded by proprietary standards and a whole lot of NP-Hard algorithms. We're working on it...

Other companies have already been working on using standard AMD servers for reconfigurable computing. Using a CPU socket is a step forward but this has been available for years using FPGAs on PCI cards.

http://www.xtremedatainc.com/xd1000_brief.html
http://www.drccomputer.com/index.html

Certainly all things have limitations. But, catch this:
1.) The hardware design (schematics, layouts, etc) are OPEN;
2.) The FPGA Verilog code is OPEN;
3.) The software is GPL.

As to the computational power of the CPU, I'm thinking an FPGA coprocessor could be used to great effect; something like a DRC coprocessor in a socket 940 (Opteron socket): see http://www.drccomputer.com/pages/modules.html for details. Run the correlation and other functions in the FPGA and offload the grunt work of the algorithm to the hardware logic you blow into the FPGA.

In my own experience, a continuum analysis (power spectrum integration using cascaded FIR filters) and a simultaneous FFT can run with 65% of a 3GHz Xeon, with all the X11 overhead taking 50% of a second 3GHz Xeon. The hard part is sustaining continuous 32MB/s writes over a period of hours (I have a Dell 2850 here with hardware RAID that can do 150MB/s writes in theory; in practice even that can skip samples). And that is using the GNUradio Python framework; tuned C would likely be less taxing on the CPU.

In contrast, we are working on other projects that are running an order or two of magnitude higher sample rates; one with be sampling 12 channels at 1.5GS/s 8 bits and performing a correlation for probing the interstellar medium using compact extragalactic sources at 2.1 and 8.5GHz. That will require the equivalent of an 800GHz Xeon; only hardware FPGA correlation is anywhere close to fast enough, and even then we're talking $10K high end Xilinx Virtex 4's.

Coprocessing FPGA's are basically required for real-time processing of this sort.

Other Processors could refer to something like this FPGA module from DRC.
An FPGA connected directly to a HyperTransport link would provide a processor that can be specified to each individual program.

Plugging a FPGA directly into the Hypertransport bus on a multisocket mobo sounds like a fantastic idea. It's not quite as easy to scale as coprocessors that live on PCI cards, but the bandwidth benefit should be huge. The downer is that these chips cost $4500 now, so the performance improvement would have to be pretty tremendous to be cost effective.

... we'll have custom HyperTransport socket FPGA chips to boost Opteron systems coming out of http://www.drccomputer.com/ real soon.

when you can just read about it on the company's website?

http://www.drccomputer.com/pages/products.html

That that this is directly related... but it is interesting and related in the sense that my first effort in DSP work was moer or less bottlenecked at the ISA bus... and lately have been tinkering with a design that certainly would be by a PCI-X or PCI-e bus.

  http://www.drccomputer.com/pages/products.html