4 ms·
This does a good job of capturing the first part of the problem. Note that Xilinx did make an "open" FPGA part, the 3000 series, it did not do well for them. T
by ChuckMcM 12y ago
This does a good job of capturing the first part of the problem. Note that Xilinx did make an "open" FPGA part, the 3000 series, it did not do well for them.
The critical thing to understand, is that FPGA users (the big ones, not the casual ones) don't want them to be open. They want the information they program into them to be secret so that their hardware is not easily duplicated by the folks in China and mass produced to under cut them. You can't patent schematics. So these folks want to have a way that a board can be assembled in China but not reproduced there in a way that isn't an exact clone (which you can get blocked as a counterfeit product).
Ok, so where does that leave us? Well it might be easier to create your own FPGA design, have TSMC make it on their cheapest process. And then try to sell those chips.
And if you're saying "Jeebus Chuck! That isn't 'easy' at all." you would be right. And that is why a 200Mhz general purpose CPU is easier to turn into a "custom chip" than an actual custom chip.
So where does that leave us? Well, all the bits between HDL and chip can be done, in a low scale way, either in simulation or on bulk small scale hardware. You can program CPLDs to be simple Logic Units (LU) for your FPGA and wire them together on a PC board. You can find the things that need to be parameterized (intra-LU timing, global clocks, I/O configuration) and build tools which can synthesize, place & route, and download to your "pseudo" FPGA. And if you have all that tooling, and you can show it works, you might be able to get a partner to make some small scale parts for you (200 - 1000 LUs)
It won't happen over night, that is a 5 year plan minimum I think. And you have to get over the hump of developing tooling for an FPGA that may never exist. On the plus side it should be good for several Masters level projects and probably a PhD or two, and if your simulation work was robust you might become the 'standard model' for testing assumptions about reducibility of different constructs into hardware[1]
[1] Sort of the Lena[2] for FPGAs.
[2] A playmate digitized by USC which became the standard image processing image for research.
- wolfgke 12y ago> Note that Xilinx did make an "open" FPGA part, the 3000 series, it did not do well for them. Could you provide further details?
- ChuckMcM 12y agohttp://www.xilinx.com/support/documentation/data_sheets/3000.pdf http://www.xilinx.com/support/documentation/data_sheets/3000... is the data sheet. I used to have the programming and layout specs doc they gave out.
- wolfgke 12y agoBut in how far didn't it do well for Xilinx?
- ChuckMcM 12y agoWell the guy I know who used to be in charge of strategic marketing for Xilinx shared that it had not met sales projections and support costs were higher (which cut into margin) when people were trying to program the things with their own tools. So sharing all that stuff about the chip alienated some big customers, reduced profit margins, and didn't help volumes. That pretty much defines 'fail' in the chip business.
- diamondman 12y agoThis is an interesting story. An unfortunate end. Most of these things sound like short term issues that get resolved as the tools are made. Alternately Xilinx could have just put the docs out there and said they only support ISE. That way if an open source developer makes a tool they have no responsibility, but can reconsider in a few years in case it gets wide acceptance. I am not a market strategist, but supporting everyone's custom tool sounds like a quick way to waste money and the wrong way to try and open source something. Similar to respecting every coupon people bring in, even hand written ones.
- ReversedGif 12y agoCan you find the "programming and layout specs docs," online or physical copies? I can't find any mention of them anywhere. I'm very curious to see them.
- yuubi 12y ago
- MichaelGG 12y agoIf there's so much value there, how are these companies stopping internal espionage? If a company wanted to see what their competitors were doing to gain a large advantage, they could certainly field a candidate to work as an agent for their competitor. It sounds like how Nikon doesn't document the NEF format and pretends their binaries are the only way to decode such files. Except obviously the secrets these hardware companies have is far more than a simple RAW format.
- ChuckMcM 12y agoFrom what I've read of internal leaks, the leakers get caught and charged and dealt with, which discourages other leakers perhaps. But there isn't anything stopping an employee from selling or stealing company secrets. That hasn't changed. But during the 70's and 80's a lot of manufacturers stopped printing schematics in response to clones appearing which used a circuit that was close (but not exactly like) theirs and sold at a deep discount[1]. Early computer board manufacturers used programmable logic to create special "nodes" within the system which obscured the functioning of things such that to clone them you would have to hire an engineer to understand the design and re-design it. Generally that was more effort than most of the cloner companies were willing to go through. (apparently if they had their own engineers that could understand the circuit they didn't need to clone, they could design their own) FPGAs get you at least two layers of obfuscation, both the bit stream is encrypted and "decompiling" from the place and route configuration bits is not straight forward at all. Folks who have gone down the path of trying to figure this stuff out (as the OP is possibly doing) discover that changing a routing bit way in the front changes a bunch of other bits later but in very subtle ways. And routing in an FPGA is a great way of screwing up function, when you do timing analysis if you have to run at 200Mhz you have to make sure your setup and hold times are accurate to within a few nanoseconds across temperature and voltage differences. FPGA vendors tout this as a 'design security' feature some times. Decoding this data is a lot more challenging than reverse engineering an image format. [1] And yes you can argue this is just business, and it is, and this is just "business defense."
- cushychicken 12y agoThanks for all these great answers you've given - they're a nice compliment to the article, really explain a lot of why these sorts of decisions were made by the FPGA companies.
- phkahler 12y ago>> They want the information they program into them to be secret so that their hardware is not easily duplicated by the folks in China and mass produced to under cut them. So how does having a closed tool chain achieve that? Aren't the FPGAs programmed either on the line, or from ROM at power up? Either way, board producers should be able to replicate the product. How does being closed help?
- fpgaminer 12y ago> So how does having a closed tool chain achieve that? It doesn't. It slows reverse engineering, but doesn't itself prevent cloning. I'm not sure what ChuckMcM was trying to argue. I don't know of any big companies that care whether or not the toolchain/devices are open source. They just want a way to get their design into the world. What we (users of FPGAs) care about is price; the toolchains are sufficient (though yes, cruddy in various ways). And to be honest, this conversation about the FPGA (and ASIC) toolchains being closed sources crops up again and again. They're closed source not because Altera/Xilinx/Microsemi need to keep secrets. They already know eachother's "secrets", and the FPGAs themselves are rather trivial devices. They're closed source not because users want them to be; again they couldn't care less. They're closed source because the userbase is absolutely tiny. A small userbase means we will only ever see tools that are sufficient. There is no benefit to Altera/Xilinx/Microsemi to put in extra time and effort open sourcing their work. And honestly, I wouldn't want them to. I'd rather their time and resources go into continuing to drive the price down and features up (as does every other company). Not that I don't think opensource toolchains/devices wouldn't be great. I follow the MiGen/Milkymist mailing list where some open source FPGA tools are being developed. But in the commercial space, open sourcing FPGAs and their tools isn't the highest priority. Going back to protecting designs, that's what bitstream encryption is for, which all modern FPGAs possess. Though Altera was quite late to the ballgame on that front with their Cyclone series...
- minthd 12y agoSure the fpga companies don't want anybody to know how to build their software. Building good fpga's is not that hard(if you're intel, or quallcom or any other of the big guys). But building good, efficient software is very hard, takes decades and tens of thousands of customers designs to run it on.