Post by thread:[EE]: FPGA computers

[OT] but plenty of FPGA stuff floats across the list. :-) (full article at link below) http://www.forbes.com/2003/03/25/cz_dl_0325star2.html Computer Hardware & Software Super-Cheap Supercomputing? Daniel Lyons, 03.25.03, 12:00 PM ET NEW YORK - Star Bridge Systems employs 16 people and rents office space in a modest beige building near a Celtic memorabilia shop and a health food store on Main Street in tiny Midvale, Utah, which isn't exactly the middle of nowhere, but close enough. The company's founder and chief technologist, Kent Gilson, is a 37-year-old high-school dropout who for years has been derided by computer scientists as, well, a bit of a fringe character. The company's chief executive, Daniel Oswald, joined Star Bridge four months ago after running a foundation that dealt with ancient religious texts and Mormon studies. In other words, IBM (nyse: IBM - news - people ) this ain't. Nevertheless, this band of outsiders claims to have created a reconfigurable "hypercomputer" that performs like a supercomputer but sits on a desktop, uses very little electricity, needs no special cooling systems and costs as little as $175,000. The secret is in the chips. Instead of yoking together hundreds or even thousands of microprocessors--as traditional supercomputers do--Star Bridge uses a dozen or so relatively inexpensive field-programmable gate array (FPGA) chips. Each FPGA can handle thousands of tasks at the same time, in parallel. A microprocessor can only do one thing at a time. So even though a traditional supercomputer might have hundreds of microprocessors, a machine with only a handful of FPGAs can outperform it. <snip> -- http://www.piclist.com#nomail Going offline? Don't AutoReply us! email spam_OUTlistservTakeThisOuTmitvma.mit.edu with SET PICList DIGEST in the body

Charles Craft wrote: >[OT] but plenty of FPGA stuff floats across the list. :-) >(full article at link below) > >http://www.forbes.com/2003/03/25/cz_dl_0325star2.html > >Computer Hardware & Software >Super-Cheap Supercomputing? >Daniel Lyons, 03.25.03, 12:00 PM ET > > Ah, but the devil is in the details. Sure, you can essentially take a computer program convert it to a logic schematic, and it will outperform in many ways the computers we use today, even supercomputers for certian tasks. For instance, if you want a machine that will simply perform floating point additions, you can fit thousands of those onto one large FPGA. Clock it fast and it will outperform many supercomputers in the FLOPs measurement. The problem is efficiently converting a program (written in C, Fortran, etc) into such a logic design. A simple program is trivial. One which simulates weather patterns over the western hemisphere and uses matrix calculations on matrices which are larger than the machine can hold inside one FPGA at once is non-trivial. Converting it into a logic pipeline which can operate at more than a modest speed ( > 100MHz) is even harder. It's something people have been working on for some time, and in a way it's a holy grail for computing, at least in large systems. If one could get an FPGA that is complex enough to simulate even a pentium II at 300MHz one could recompile programs to run much faster than todays multi-gigahertz systems. It wouldn't do well for desktops, as people like to run several different programs at once, and the task switching of re-configuring the fpga for each task would kill the performance gain, even if you had enough room to put 3-4 programs on the FPGA at once. Some cheap researchers in the past purchased broken memory and processors, wired them up with FPGAs, and created a control unit that routed around bad memory cells, and only allowed operations to be performed on processers which could handle a specific operation (ie, one CPU might have a bad FPU, so they disallowed FPU type operations on it). Neat for research, but probably isn't worth the work. -Adam -- http://www.piclist.com#nomail Going offline? Don't AutoReply us! email listservKILLspammitvma.mit.edu with SET PICList DIGEST in the body

Hi Jan-erik! On Wed, 26 Mar 2003, Jan-erik Söderholm (QAC) wrote: > A couple of Swedish guys are doing this. > > Read more : http://www.mitrion.com/technology.shtml -Adam's comments were essentially corrrect, "It wouldn't do well for desktops, as people like to run several different programs at once, and the task switching of re-configuring the fpga for each task would kill the performance gain, even if you had enough room to put 3-4 programs on the FPGA at once." As I was reading that I was thinking of exactly the kind of things that the Mitrion page described. The intention is more of an 'application on a chip' thing. Yes, it takes a while to reprogram so can't flip between various apps very quickly within a given chip. But an accountant for instance doesn't do that, he has a set of specific apps and that's all that he uses. Maybe a plugin board with half a dozen FPGAs for different apps. One each for PIC compiler, simulator, and board layout. Instead of running programs all the PC CPU does is supervise the FPGAs and manage I/O. Have a :) day! jb -- jim barchuk EraseMEjbspam_OUTTakeThisOuTjbarchuk.com -- http://www.piclist.com hint: To leave the PICList piclist-unsubscribe-requestspam_OUTmitvma.mit.edu>

It would be appropriate for high end graphics programs and video rendering for instance. Possibly like projects along the lines of the renderwall used to animate some feature films and special effects add-ins which normally take a looooooooong time on some 'fairly serious' clusters. But even a lot of home and office workstations would benefit on a more volume oriented approach to graphics processing, this is often a painfully slow exercise.( though it would be considered nothing in comparison to projects like they use the renderwall on) JJ > {Original Message removed}

OK...had to read thru all the comments and then add my own. First off....there are some correct assumptions...doing things in hardware, specific tasks for instance that are the ONLY task it does, as in graphics rendering do VERY well in hardware. I've done it in my past life. Second, using them as a pre or co processor also does well....IF....again, its doing a specific task only. And it still applies to cluster computers as well. Now, back to Starbridge. It's based on Xilinx FPGA's, can't recall what part exactly. They have been around...TRYING to push this thing for the last 8 years or so. Just recently they had a press release saying they sold 5 units to the gov't (nice way to spend our tax dollars in the states). Yet....does it work? Jury is still out and probably will be for some time. And yes....they are located not far from me, so I do have some insight on the company and have talked to the Xilinx FAE in passing about it...and he is very familiar with this. I won't say anymore....least not on the forum. _________________________________________________________________ -- http://www.piclist.com hint: To leave the PICList @spam@piclist-unsubscribe-requestKILLspammitvma.mit.edu>

I can see them being extremely useful for cryptographic work. It's the ability to prototype hardware, but for computer scientists who don't want to deal with the actual implementation. -Adam Micro Eng wrote: {Quote hidden}> OK...had to read thru all the comments and then add my own. > > First off....there are some correct assumptions...doing things in > hardware, > specific tasks for instance that are the ONLY task it does, as in > graphics > rendering do VERY well in hardware. I've done it in my past life. > > Second, using them as a pre or co processor also does > well....IF....again, > its doing a specific task only. And it still applies to cluster > computers > as well. > > Now, back to Starbridge. It's based on Xilinx FPGA's, can't recall what > part exactly. They have been around...TRYING to push this thing for the > last 8 years or so. Just recently they had a press release saying > they sold > 5 units to the gov't (nice way to spend our tax dollars in the states). > Yet....does it work? Jury is still out and probably will be for some > time. > > And yes....they are located not far from me, so I do have some insight on > the company and have talked to the Xilinx FAE in passing about > it...and he > is very familiar with this. I won't say anymore....least not on the > forum. > > > > > _________________________________________________________________ > > -- > http://www.piclist.com hint: To leave the PICList > KILLspampiclist-unsubscribe-requestKILLspammitvma.mit.edu > > > -- http://www.piclist.com hint: To leave the PICList RemoveMEpiclist-unsubscribe-requestTakeThisOuTmitvma.mit.edu>

I'm wondering if they'd be good for a neural network. It sounds like I could put a "trained" neural network onto an FPGA. Would they be better than a DSP for specialized vision processing (as in preprocessing and image for robotics)? {Quote hidden}> I can see them being extremely useful for cryptographic work. It's the > ability to prototype hardware, but for computer scientists who don't > want to deal with the actual implementation. > > -Adam > > Micro Eng wrote: > > > OK...had to read thru all the comments and then add my own. > > > > First off....there are some correct assumptions...doing things in > > hardware, > > specific tasks for instance that are the ONLY task it does, as in > > graphics > > rendering do VERY well in hardware. I've done it in my past life. > > > > Second, using them as a pre or co processor also does > > well....IF....again, > > its doing a specific task only. And it still applies to cluster > > computers > > as well. -- D. Jay Newman ! Pudge controls the weather. spamBeGonejayspamBeGonesprucegrove.com ! http://enerd.ws/~jay/ ! Oh good. My dog found the chainsaw. -- http://www.piclist.com hint: To leave the PICList TakeThisOuTpiclist-unsubscribe-requestEraseMEspam_OUTmitvma.mit.edu>

Vision processing would suit it perfectly, however, neural networks,IIRC even trained ones still need to vary the program,(if I'm wrong please somebody correct me) although it is less of a major change after the training is done, it is still an adaptive system, machine vision , image recognition, encryption and crypto cracking etc are all computationally intensive and the program doesn't change and thus, fit perfectly. > {Original Message removed}

There is no problem implementing a self-modifying FPGA, but you either need to design it so it has adjustable behaviour or select a model where you can alter the programming-bits from within the device. KreAture ----- Original Message ----- From: "Jonathan Johnson" <RemoveMEjonathanTakeThisOuTOUTEREDGE.NET> To: <PICLISTEraseME.....MITVMA.MIT.EDU> Sent: Friday, March 28, 2003 3:00 AM Subject: Re: [EE]: FPGA computers > Vision processing would suit it perfectly, however, neural networks,IIRC > even trained ones still need to vary the program,(if I'm wrong please > somebody correct me) although it is less of a major change after the > training is done, it is still an adaptive system, machine vision , image > recognition, encryption and crypto cracking etc are all computationally > intensive and the program doesn't change and thus, fit perfectly. > > > {Original Message removed}