1

Assuming there is a task (e.g. an image processing method with a lot math) which is reasonable to be implemented on FPGA in sense of answer https://stackoverflow.com/a/8695228/544463

Is there any known (that you can actually name) successful application or practice for combining it with "dedicated" (designed on custom demand) super computing cluster (HPC), e.g. with Infiniband stack? I wonder if that has already been done and to which extend that was successful.

My main motivation for the question is that http://en.wikipedia.org/wiki/Reconfigurable_computing is a long term (academic) perspective for the future development of cluster computing as a distinctive alternative to cloud computing (the later concentrates more on the software (higher) flexibility level but also through possible "reconfiguration"). Is it already practical?

I would also expect somebody is doing research on this... It would be nice to learn about results.

Community
  • 1
  • 1
Yauhen Yakimovich
  • 13,635
  • 8
  • 60
  • 67
  • Please note that in terms of http://stackoverflow.com/questions/317731/cuda-vs-fpga a valid answer could sound like "No, it is not practical" or "FPGA is not very flexible/complicated to work with", but I am looking more for an actual experience of doing so being shared and/or doing research on reconfigurable hardware architecture scalable to cluster computing in broader sense (i.e. common-sense and hand-waving arguments are obviously speaking against such undertaken, but people can be stubborn enough to work on this anyway). – Yauhen Yakimovich Mar 28 '12 at 11:47
  • I kindly thank you all for your answers. Have to study each of them more thoroughly. But already at first look - using FPGA for HPC seems very viable in both short and longer terms – Yauhen Yakimovich Apr 03 '12 at 18:33

8 Answers8

3

Well, it's not FPGA, but D.E. Shaw's Anton computer for molecular dynamics is famously ASICs connected with a custom high-speed network; J. P. Morgan uses clusters of FPGAs in its risk-analysis calculations (recent Forbes article here). Convey computers has been pushing FPGA+x86+high speed networking fairly hard for the past couple of years, so presumably there's some sort of market there...

Jonathan Dursi
  • 50,107
  • 9
  • 127
  • 158
3

http://www.maxeler.com/ - they build racks of Intel PCs hosting custom boards stuffed with FPGAs (and - critically - the associated software and FPGA code) to speed up seismic processing, financial analysis and the like.

I think they could be regarded as successful (I gather they turn a profit) and have big customers from finance and oil companies amongst their clientele.

Martin Thompson
  • 16,395
  • 1
  • 38
  • 56
2

Some have already been mentioned (convey, cray), some not (e.g. beecube). But one of the biggest FPGA-Clusters I ever heard of, is missing:

The Large Hadron Collider at CERN. They produce in seconds enormous amounts of data (2.7 Terabit/s). They use the FPGAs (> 100) of them to reduce and filter the data to reduce it, and make it handable.

It does not fit your request to be connected to a dedicated HPC-Cluster, but they are a HPC-Cluster on their own (as on the higher hierarchy levels the used FPGAs are FX, they include two PowerPCs and are also some kind of "normal" cluster).

flolo
  • 15,148
  • 4
  • 32
  • 57
2

I know that Cray used to have a series of supercomputers some years ago that combined AMD Opterons with Xilinx FPGAs (iirc) through a HyperTransport bus, basically allowing you to create your own specialized processor for custom workloads. According to their website though, they now seem to have dropped FPGAs in favor of GPUs.

For the current research, there's always Google Scholar...

Update: After a bit of searching, it appears to have been the Cray XT5h, which had the possibility of using FPGA coprocessors...

sonicwave
  • 5,952
  • 2
  • 33
  • 49
2

You might be interested in Maxwell.

High Performance Mark
  • 77,191
  • 7
  • 105
  • 161
2

Is there any known (that you can actually name) successful application or practice for combining it with "dedicated" (designed on custom demand) super computing cluster (HPC), e.g. with Infiniband stack? I wonder if that has already been done and to which extend that was successful.

It's being attempted academically with Novo-G.

1

There is quite a lot of published work in reconfigurable computing applications.

Here's a list of links to SRC Computers-centric published papers.

There's the Center for High-Performance Reconfigurable Computing.

Google search "FPGA" or "reconfigurable" along with these academic institution names and you'll find many published papers. Some of the papers you'll find go back to 2004.

Jackson State University
Clemson University
Catholic University
George Washington University
George Mason University
National Center for Supercomputing Applications (NCSA)
University of Illinois (UIUC)
Naval Postgraduate School (NPS)
Air Force Research Lab (AFRL)
University of Dayton Research Institute (UDRI)
University of Florida
University of Arkansas

There also was a reconfigurable-centric conference hosted by NCSA, the Reconfigurable Systems Summer Institute (RSSI).

This list is certainly not exhaustive, but it will get you started.

Disclosures: I currently work for SRC Computers, LLC, I worked at NCSA/UIUC and I chaired the RSSI conference its first two years.

David Pointer
  • 905
  • 2
  • 18
  • 31
0

Yet another great use case developed by adapteva called parallela (they have a kickstarter project). They are developing a epoch-series of processors controlled by a two cores ARM processor (that shares the board).

epoch outline

I am so much anticipating to have this toy in my hands!

PS

Since it was largely inspired by ardunio (and similar ARM-like) systems, this project is still limited by 1 Gbps networking.

Yauhen Yakimovich
  • 13,635
  • 8
  • 60
  • 67