EHLO, board designers

Why can I not purchase a PCI board with a Spartan3, some SRAM, and a DRAM slot for $200? That's way more than the cost of the parts. Heck, with Xilinx's recent anouncements about 3rd party PCI-Express support I should be able to get that interface for the same price. What I'm picturing is general coprocessing. Nobody seems to think FPGAs are valuable for everyday coprocessing if you judge by the boards being made. Cray, Starbridge, SGI, Nallatech, and others recognize the value in coprocessing but they are targeting the HPCS market. It's the wrong market. All the other PCI boards I've seen lately are made for prototyping or DSP processing, especially in the sub $5k range. A simple Spartan3-based board for coprocessing could change the world. Video game companies could ship neural net modules, math companies could ship libraries that use it for acceleration, CAD and imaging companies could take advantage of it for acceleration, etc; for that to happen, everybody needs one. For that to happen, they have to be as cheap as a decent graphics card and compilers for them must be as cheap as GNU gcc. And they don't need a freakin' serial port, ethernet port, parallel port, USB port, half a dozen different RAM ports, proprietary connectors, and the kitchen sink for prototyping! How are we going to get there and why are we not there already?

Reply to
Brannon
Loading thread data ...

"Brannon" schrieb im Newsbeitrag news: snipped-for-privacy@g47g2000cwa.googlegroups.com...

changing the world huh? challenging is it not?

the board you wish isnt there, either somethign missing or something too much.

I would say it makes MUCH more sense to get it done with PCIe - hmmm as my last inquiry the Avnets Spartan-3 + PCIe evaluation kit was supposed to be available end of June - but I havent asked about it for some while maybe its even out already.

there are some cheap Lattice EC boards with DIMM sockets, but also a little above the 200$ range

humm the PCIe board could actually be CHEAPER than the PCI board - smaller and no need for those 5V translators :)

Antti

Reply to
Antti Lukats

Hi Antti,

True...

... but I'd first have to shell out $500 for a new mobo, CPU and memory. Hmmm... how to get budget for that...

Ben

Reply to
Ben Twijnstra

"Brannon" wrote in message news: snipped-for-privacy@g47g2000cwa.googlegroups.com...

Because, given the low demands for boards like that, it costs a lot more than $200 once you factor in the time needed for someone to design & debug that board. You might expect Xilinx to do this -- many semiconductor sell evaluation boards at what's effectively "giveaway" prices in that they lose money on the sale, but figure they'll make it up in sales of the actual part -- but unfortunately Xilinx doesn't (at present).

Costs of parts is an incredibly small factor in determining sales price in many products.

You need a lot more than a simple evaluation board to get all those companies to adopt your co-processing model -- you also need large, sophisticated libraries that help programmers make use of the co-processing features. And keep in mind some of the main downfalls of any hardware based co-processor: It's a given that Intel, AMD, etc. will crack out faster and faster processors every year, so your specialized hardware ends up having to be re-designed regularly to continue to keep its "edge" -- this is an on-going expense (Patterson & Hennessy, in their Computer Architecture book, talk about how IBM once had a hard drive with a specialized co-processor to perform text string searches, but that over time as CPUs became faster and the specialized hardware remained the same, eventually it became slower to use that specialized hardware than to just burn CPUs cycles to perform the search. Similarly, by the time the Commodore Amiga computers were running 25MHz 68030 CPUs, it had become faster to use the CPU to perform some graphical functions such as text scrolling than to use the old Blitter graphics co-processor running at 7.14MHz). If your application really DOES catch on, it'll just be designed into a custom chip or made part of a CPU anyway -- witness Intel's MMX instructions and the proponderence of WinModems today, look at how graphics cards now have "GPUs" that are often on par in sophistication with general purpose CPUs, notice how the better sound cards have on-board ASICs containing lots of DSP hardware, etc.

So... all I'm really saying is that, while there certainly are applications for "general purpose" co-processing, any application of such technology that's particularly successful will just be moved into an ASIC where far more people will benefit from lower price, lower power consumption, etc. anyway.

How about if you put in your own time and effort to design such a card and its corresponding software (as an "open source" hardware project) and they start selling the boards for $200? Many people here would probably buy one! :-)

---Joel Kolstad

Reply to
Joel Kolstad

The problem is that the number of applications usefull for normal users where a "tiny" FPGA like Spartan3 can make a significant speed-up in comparision to a 3.8GHz Pentium IV, is quite small. Do you notice that currently, GPUs for example are monsters with 300+ million transistors? FPGAs, even the fattest ones are way too small to justify their application in GENERAL processing markets. Don't take me wrong, I do DSP in FPGA and not with processors exactly for performance reasons but I use it for very special applications not for say, accelerating games or CAD applications as you suggest. If you could get an FPGA board with a capacity of 10 million gaits in less than $200 in the market, then I could start thinking of some small applications for it in general PC market. With smaller FPGAs, just forget it.

Reply to
Arash Salarian

Making the case as a board manufacturer the cost of a board is not just the chips on there. The development board market is not a hugh volume market and usually assembly and test of boards is considerably more that you would incur in a consumer product volume level. Companies in this market are also not charities and at least like to get the development labour costs back on a board development. Man time, or even women time, is expensive and if advertised over a few hundred boards still makes a considerable effect on a board price. And to cap all that it is even nice to make the odd bit of profit. This tends to keep the shareholders happy which is often a good thing.

And to make our case I agree there are a lot boards out there than have features that most people individually don't want but you don't want to make individual boards for everyone due to the costs involved. As we don't have manufacturer lines to push, unlike some of our competitors, our approach here is to go for a minimalist fixed fit of functions and then to make up for this with cheap simply manufactured modules. We also design our boards so that you can add your own add-ons easily.

Our boards come close to general use scenario and we already supplying some number of boards in this way. Volume customers do get substantial discounts as the batch handling, shipping etc has a much lower cost base for us to cover.

At the moment PCI-E is a lot dearer to implement than 32bit PCI. The PCI-E core also eats a large part of your FPGA unless you have a large device to start with. The Philips device listed for the phy part of the design is currently hard to get, and dearer than the 3 bus switches we use on 32 bit PCI, although I believe that supply issue will ease shortly.

And to end this long rant and to wake up those skinflints now sleeping we will have a cheap product that at least in part will meet your needs late September or early October. Still not happy then go make your own.

John Adair Enterpoint Ltd. - Soon to be the home of Raggedstone1. The very very cheap PCI Development Board.

formatting link

Reply to
John Adair

I disagree, there are quite a bit of applications where a Spartan3 can make a significant speed-up in comparison to a 3.8 GHz P4. These designs, however, are generally not trivial and take a considerable amount of expertise in the generally non-overlapping disciplines of FPGA design and algorithm design. The upper end of the spartan3 line is bigger, faster, and much more capable than the biggest FPGAs were a couple years ago (which I successfully used in many DSP applications that beat the pants off the software equivalents, thank you very much). The issue is not the capability of the 'cheap' FPGAs, rather it is the complexity of the design required to make efficient use of the parts in computationally intensive applications. Tools for hardware DSP are still in their infancy, and the combined expertise for an efficient design that will fit these parts and still perform well is exceptionally rare. Bottom line is programming FPGAs is not like driving your father's CPU:the design has many more degrees of freedom, which makes it harder both for a designer as well as for automated tools to arrive at the 'right' solution. A high volume low cost FPGA accelerator for garden variety PCs isn't going to happen until the automated tools make hardware DSP as easy as computer programming. Until that happens, there just isn't going be enough demand to amortize the development and support costs over enough units to make the cost of the components the major cost of the board.

--
--Ray Andraka, P.E.
President, the Andraka Consulting Group, Inc.
 Click to see the full signature
Reply to
Ray Andraka

ElectronDepot website is not affiliated with any of the manufacturers or service providers discussed here. All logos and trade names are the property of their respective owners.