Forum Discussion

Altera_Forum's avatar
Altera_Forum
Icon for Honored Contributor rankHonored Contributor
19 years ago

Altera FPGAs in High Performance Computing(HPC) Applications

I am trying to better understand the different methodologies on using Altera FPGAs as a coprocessor in HPC applications.

Believe there there are four known methodologies. Familair with the first three..

1. Module plugs into processor socket connection-Intel FSB

2. Module plugs into processor socket connection-AMD HyperTransport

3. Module plugged via PCI express interface

4. a. Module plugs into system memory DIMM socket

b. Module plugs into a hard drive bay and connects to processor via DIMM socket

Trying to get a better feel for Item-4. Wanted to know the pro's and con's of using FPGA on a memory DIMM socket?

Does anybody know of any third party companies that use this approach and offer solutions in this space?

Thanks!

4 Replies

  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor

    SRC Computers is the company I've heard of for# 4. They were started by Seymour Cray and do a number of different accelerator-type products. Also be sure to investigate the development side, i.e. hardware/software interaction. If you're ready to write HDL, it's usually not a big deal, but a lot of users investigating this area only write C for processors and want to "push a button" to create hardware from their C.

    www.srccomp.com
  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor

    Yes, I did look into Xtremedata products for# 1 and# 2. And for# 3 found GiDEL's solutions suitable www.gidel.com. They have have a tool that can map DSP Builder model based designs to Altera FPGA via PCI express interface(Hardware In the Loop Methodology).

    BTW: I am interested in seeing data throughput MB/s for two of Xtremedata's products

    1. 2S180 FPGA connected via Intel FSB to a Xeon processor(Xtremedata)

    FSB interface supports 1066MHz and 64 bits. FPGA I/O connected via FSB might have some speed limitations. What I am looking for the real data throughput that was achieved b/w a 2S180 and a Xeon processor(MB/s)?

    2. 2S180 FPGA connected via AMD HyperTransport to an Opteron Processor

    HyperTransport standard supports 3.2GB/sec(1.6 each way). Again the data I am looking for is the actual throughput between a 2S180 and an Opteron processor.

    Any thoughts on this?

    Thanks!