Custom Search
Site Home Archive Home FAQ Home How to search the Archive How to Navigate the Archive
Compare FPGA features and resources
Threads starting:
Authors:A B C D E F G H I J K L M N O P Q R S T U V W X Y Z
On Thu, 28 Oct 1999 17:23:00 -0400, Ray Andraka <randraka@ids.net> wrote:
>Oops... I was about to yell back to say I didn't say altera was better in all
>other apps, but reading my post that is what my fingers typed. I would by no
>means say that Altera is best for everything else---it just plain isn't. Those of
>you who know me know that what I meant to type was "For other applications, the
>Altera is sometimes better". Altera tends to do better at collecting "random"
>logic, especially for the user who isn't willing to do placement. (Again, this is
>not meant to be a blanket statement by any means). Xilinx is better in the vast
>majority of datapath designs, and is a hands down winner in arithmetic designs or
>designs requiring delay queues. When it comes down to it, the best device depends
>on the application, and might even include non-technical considerations such as
>tools ownership, designer experience etc.
For the "random" logic performance issue, doesn't it merely related to
the synthesis tool??
Child :~{}
Article: 18551
NO.
Assuming reasonable synthesis, the routing delays and ability to handle wide fanins
have a large impact on performance. Altera's routing delays are relatively independent
of placement, so having not-so-local routing does not degrade performance
significantly. Connections within a LAB in Altera are quite fast, and the altera
compiler does a fairly good job at keeping stuff within a LAB when the stuff is
reasonably small (provided carry chains don't prevent that). In xilinx, routing delay
is much more dependent on placement, and the auto-place tools really don't do a very
good job. Local routing in the xilinx is very fast, and very structured. As you move
away from the structured placement there is a significant degradation of maximum clock
rate. Random logic takes longer to hand place than a structured data path.
Child K.L. Sun wrote:
> For the "random" logic performance issue, doesn't it merely related to
> the synthesis tool??
>
> Child :~{}
--
-Ray Andraka, P.E.
President, the Andraka Consulting Group, Inc.
401/884-7930 Fax 401/884-7950
email randraka@ids.net
http://users.ids.net/~randraka
Article: 18552This is a multi-part message in MIME format. --------------BB51596145F9BE5180CC1A9E Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Hi, Arrigo Benedetti wrote: > does anyone have the reference to the paper where the FPGA > was first introduced? > The question is bit more complex than it sounds: - as far as I know the term FPGA was created by Actel's marketing people but there were many programmable architectures around before the term was coined. - assuming you take the essential requirements of 'FPGA' as being general purpose routing (not PAL like arrays) and electrical field reprogrammability (i.e. not patch panels, laser cutting of links etc) then I'd recommend reading the following references and making your own mind up: S.E. Wahlstrom 'Programmable Logic Arrays' Electronics Vol 40, No 25, Dec. 11 1967. R.C. Minnick 'Survey of Microcellular Research', J. ACM, Vol 14, No. 2, April 1967. Note, before I get flamed I am not claiming that either of these are definitively the first paper about FPGA's. The first document on the modern FPGA is probably the famous Freeman patent US 4,870,302 (re-issued as Re. 34,363). Tom Kean. Tom > Thanks in advance, > > -Arrigo > -- > Dr. Arrigo Benedetti e-mail: arrigo@vision.caltech.edu > Caltech, MS 136-93 phone: (626) 395-3695 > Pasadena, CA 91125 fax: (626) 795-8649 --------------BB51596145F9BE5180CC1A9E Content-Type: text/x-vcard; charset=us-ascii; name="tom.vcf" Content-Transfer-Encoding: 7bit Content-Description: Card for Tom Kean Content-Disposition: attachment; filename="tom.vcf" begin:vcard n:Kean;Tom tel;fax:UK +44 131 556 9247 tel;work:UK +44 131 556 9242 x-mozilla-html:TRUE org:Algotronix Ltd. adr:;;P.O. Box 23116;Edinburgh;;EH8 8YB;Scotland version:2.1 email;internet:tom@algotronix.com title:Director note:Web Site: www.algotronix.com x-mozilla-cpt:;4768 fn:Tom Kean end:vcard --------------BB51596145F9BE5180CC1A9E--Article: 18553
If I can put a different point of view. I'm new to the game, but I just implemented a signal processing type design, which I wrote as 'behavioural-ish' (ie not aimed directly at being synthesisable) VHDL. It isn't huge, but it has a collection of multipliers/adders/comparators. I thought I'd just see how far off a usefully synthesizable design it was so I ran it through FPGA Express. It runs at 4MHz in a Virtex, and 14MHz in Flex10K (and 7MHz in a Flex6016 - just to see). 4MHz is just fast enough, but the Altera device gives me plenty of spare time to play with. The big plus for me is that I can get the performance I want without having to mess around hand-tuning the algorithm to the architecture, and at the moment, all I'm (or my boss is :-) looking for it proof of concept. It appears from what I've gathered so far - and as I say I am new to this - that Xilinx will perform better, given the up front investment. However, the Altera devices may well perform *well enough* straight out of the box with very 'ordinary' VHDL. So there's another view: The choice may depend on how much time/inclination you have to tweak your implementation. Which brings me to a question... This design fits in a 10K20, which is nominally a 20K gates device, so clearly its not a huge design. Am I going to find the Altera architecture limits me, and will also require hand-tuning, as the design size goes up? Martin On Thu, 28 Oct 1999 10:45:39 -0400, Ray Andraka <randraka@ids.net> wrote: >Big architectural differences. Which is best depends on your >application. For signal processing, Xilinx is a hands down winner. For >other apps, the Altera is better. Look for my previous posts on this >subject in deja-news for more details. > >Child K.L. Sun wrote: > >> Have you tried both of the chips/software from these two >> companies? >> What's the difference between them? >> >> Child > > > >-- >-Ray Andraka, P.E. >President, the Andraka Consulting Group, Inc. >401/884-7930 Fax 401/884-7950 >email randraka@ids.net >http://users.ids.net/~randraka > > Martin Thompson martin@the-thompsons.freeserve.co.uk http://www.the-thompsons.freeserve.co.uk/Article: 18554
At only 4 MHz, you are leaving alot of the FPGA's potential performance on the table. If your sample rate only needs 4 MHz, you ought to consider bit serial design so that you can get into a smaller and cheaper part, for example the Xilinx spartan series. Bit serial hardware is typically about 1/nth the hardware for an n bit wide process, although it also requires an n times bit clock. you can use a hybrid digit-serial approach if the clock is too high. Altera does tend to do better with uninspired synthesis. That is mostly a consequence of the global routing structure which makes placement considerably less critical than xilinx. However, that is a double edged sword: that same global routing limits how fast you can make a design go if you are willing to put in the extra effort. For signal processing applications, Altera is handicapped by the carry arithmetic structure which limits you to a two input arithmetic function in one level of logic (compared to 3 inputs in Xilinx 4K and 4 inputs in Virtex). That means an add/subtract or accumulator with clear is forced to two levels of logic in ALtera. Also, and probably more significant, Xilinx lets you use the LUTs in the array as RAM or as shift registers, so that a 16 clock by 1 bit delay fits into the space of one LUT. ALtera has nothing similar, so you wind up using an LE for each clock delay for each bit. It gets rather expensive when you are dealing with word wide, multiple clock delay queues of the type you find frequently in filters etc. So, if your application uses delay queues or arithmetic with more than two inputs, you are going to use up alot of your logic in ALtera, where you wouldn't in xilinx. Thus, you need a device with more "marketing gates" to fulfill the same function. Also, the inability to use the LUTs as RAM is a handicap if you are doing adaptive filtering, as it keeps you out of some significant algorithmic optizations. For low data rate stuff, I would prefer to use a smaller device and go with a bit serial or digit serial design to keep unit costs down. The exception is where the higher clock rates might cause you problems. The extra design effort required for that type of architectural optimization is pretty small. martin@the-thompsons.freeserve.co.uk wrote: > If I can put a different point of view. > > I'm new to the game, but I just implemented a signal processing type > design, which I wrote as 'behavioural-ish' (ie not aimed directly at > being synthesisable) VHDL. It isn't huge, but it has a collection of > multipliers/adders/comparators. I thought I'd just see how far off a > usefully synthesizable design it was so I ran it through FPGA Express. > > It runs at 4MHz in a Virtex, and 14MHz in Flex10K (and 7MHz in a > Flex6016 - just to see). 4MHz is just fast enough, but the Altera > device gives me plenty of spare time to play with. The big plus for > me is that I can get the performance I want without having to mess > around hand-tuning the algorithm to the architecture, and at the > moment, all I'm (or my boss is :-) looking for it proof of concept. > > It appears from what I've gathered so far - and as I say I am new to > this - that Xilinx will perform better, given the up front investment. > However, the Altera devices may well perform *well enough* straight > out of the box with very 'ordinary' VHDL. > > So there's another view: The choice may depend on how much > time/inclination you have to tweak your implementation. > > Which brings me to a question... This design fits in a 10K20, which > is nominally a 20K gates device, so clearly its not a huge design. Am > I going to find the Altera architecture limits me, and will also > require hand-tuning, as the design size goes up? > -- -Ray Andraka, P.E. President, the Andraka Consulting Group, Inc. 401/884-7930 Fax 401/884-7950 email randraka@ids.net http://users.ids.net/~randrakaArticle: 18555
hi, I looked at Xilinx web page and the choices there confused me a little bit. I am trying to decide on an FPGA family for a very aggresive DSP design with Verilog. I already have synthesis and Altera P&R tools. What is the least expensive package I can use from Xilinx to P&R the output of the synthesis for Virtex(-E) and see the performance ? I don't need to program any parts yet and I don't need any design entry in the Xilinx tool. I need at least the 400K part. thanks muzo Verilog, ASIC/FPGA and NT Driver Development Consulting (remove nospam from email)Article: 18556
Suppose I had a design that was written without any particular hardware in mind. Maybe it was written by a software guy who didn't know about fan-in or routing. How much can I gain in speed or density by rewriting it to be more friendly to a paritcular architecture? Are there a tricks that an experienced designer would use from the beginning if he knew the target architecture? Is there a list of that sort of tricks someplace? Maybe a web page for each architecture? Or are they all so "obvious" that nobody bothers to write them down? How hard would it be to teach the compiler about those tricks? Would hints from the programmer/designer help the compiler? Or should it be able to figure all that out since it knows all about the target architecture? -- These are my opinions, not necessarily my employers.Article: 18557
If you missed the article in the October 18, 1999, issue of the
Electronic Engineering Times (page 73) -- you can read about the
Versatile Placement and Routing (VPR) package for FPGAs, and
download the software (free for noncommercial use) at the University
of Toronto at http://www.eecg.toronto.edu/~vaughn/vpr/vpr.html
Greg_
October 31, 1999.
{}{}{} Posted via Uncensored-News.Com, http://www.uncensored-news.com {}{}{}
{}{}{}{} Only $8.95 A Month, - The Worlds Uncensored News Source {}{}{}{}
{}{}{}{}{} Five News Servers with a BINARIES ONLY Server {}{}{}{}{}
Article: 18558On Sat, 30 Oct 1999 16:27:50 -0400, Ray Andraka <randraka@ids.net> wrote: >At only 4 MHz, you are leaving alot of the FPGA's potential performance on >the table. If your sample rate only needs 4 MHz, you ought to consider bit >serial design so that you can get into a smaller and cheaper part, for >example the Xilinx spartan series. Bit serial hardware is typically about >1/nth the hardware for an n bit wide process, although it also requires an n >times bit clock. you can use a hybrid digit-serial approach if the clock is >too high. > My point was that smaller/cheaper is not the issue for me at the moment, once smaller/cheaper becomes important, we'll be in ASIC territory anyway. >Altera does tend to do better with uninspired synthesis. That is mostly a >consequence of the global routing structure which makes placement >considerably less critical than xilinx. However, that is a double edged >sword: that same global routing limits how fast you can make a design go if >you are willing to put in the extra effort. Presumably also a double edged sword in that if you can't/don't want to put in the extra effort it makes life harder in 'Xilinx-land' <snip lots of useful comparison of architectures> > >So, if your application uses delay queues or arithmetic with more than two >inputs, you are going to use up alot of your logic in ALtera, where you >wouldn't in xilinx. Thus, you need a device with more "marketing gates" to >fulfill the same function. Also, the inability to use the LUTs as RAM is a >handicap if you are doing adaptive filtering, as it keeps you out of some >significant algorithmic optizations. > I haven't looked hard enough at the architectures to work it out, but are they not both LUT based? What extras do the Xilinx arch. have that enables all these other features? And why aren't Altera interested in copying them :) >For low data rate stuff, I would prefer to use a smaller device and go with a >bit serial or digit serial design to keep unit costs down. The exception is >where the higher clock rates might cause you problems. The extra design >effort required for that type of architectural optimization is pretty small. > Again, units costs aren't really the issue, although if I can get an order of magnitude difference in price ... :) I have to say that I am torn, in that we already use Altera devices (7K's) for a number of applications, but everything I read here implies that we need to look at Xilinx for these new applications. Most of our applications so far have been glue-logic type things rather than arithmetic based work. My other problem is that I'm not au fait enough with either the Altera 10K's or the Xilinx devices to really know how far I could push my example design if I put the effort in (without actually doing it, which will involve a significant investment in terms of time and needing to buy Xilinx tools we may not need). Anyway, thanks for the input, Martin >martin@the-thompsons.freeserve.co.uk wrote: > >> If I can put a different point of view. >> >> I'm new to the game, but I just implemented a signal processing type >> design, which I wrote as 'behavioural-ish' (ie not aimed directly at >> being synthesisable) VHDL. It isn't huge, but it has a collection of >> multipliers/adders/comparators. I thought I'd just see how far off a >> usefully synthesizable design it was so I ran it through FPGA Express. >> >> It runs at 4MHz in a Virtex, and 14MHz in Flex10K (and 7MHz in a >> Flex6016 - just to see). 4MHz is just fast enough, but the Altera >> device gives me plenty of spare time to play with. The big plus for >> me is that I can get the performance I want without having to mess >> around hand-tuning the algorithm to the architecture, and at the >> moment, all I'm (or my boss is :-) looking for it proof of concept. >> >> It appears from what I've gathered so far - and as I say I am new to >> this - that Xilinx will perform better, given the up front investment. >> However, the Altera devices may well perform *well enough* straight >> out of the box with very 'ordinary' VHDL. >> >> So there's another view: The choice may depend on how much >> time/inclination you have to tweak your implementation. >> >> Which brings me to a question... This design fits in a 10K20, which >> is nominally a 20K gates device, so clearly its not a huge design. Am >> I going to find the Altera architecture limits me, and will also >> require hand-tuning, as the design size goes up? >> > > > >-- >-Ray Andraka, P.E. >President, the Andraka Consulting Group, Inc. >401/884-7930 Fax 401/884-7950 >email randraka@ids.net >http://users.ids.net/~randraka > > Martin Thompson martin@the-thompsons.freeserve.co.uk http://www.the-thompsons.freeserve.co.uk/Article: 18559
On Sun, 31 Oct 1999 14:00:33 GMT, martin@the-thompsons.freeserve.co.uk wrote: >My other problem is that I'm not au fait enough with either the Altera >10K's or the Xilinx devices to really know how far I could push my >example design if I put the effort in (without actually doing it, >which will involve a significant investment in terms of time and >needing to buy Xilinx tools we may not need). does this mean that you haven't got any xilinx tools? if so, how are you doing your performance comparisons - are you relying on figures reported by FPGA express? if you are: 1) forget it. you can't compare two designs unless they're both routed, and the vendor's timing analyser has given you a max frequency. it's not unusual to get at least a factor of two speed difference between the synth's estimate and the real device. 2) FPGA express seems to have been slow to catch up with virtex. i believe that you should be on v3.3 to get useful results. 3) i went through the 10k v. virtex exercise with a customer who was already using 10Ks, with no xilinx experience, about 9 months ago, for a DSP application. virtex was the no-brainer then, even though the virtex tools were poor at that stage. the altera FAE didn't even bother trying to change their minds. 20K vs. virtex might be a different matter, if it ever happens. evanArticle: 18560
Andy Peters wrote: > I'm using an XC4KE part, I have a 32-bit bidirectional bus that comes out to > the pins and I'd like to constrain the tristate enable so that it's faster. > ... I have a signal called dataoe, which is a flip-flop output, and > that signal is used as the tri-state enable for the outputs, which are > called ramdata[31:0]. I have usually not had very good luck getting the Xilinx tools to improve timing in a case like this by adding timing constraints. What I would suggest is that you floorplan the dataoe flip-flop to the middle of the tristate bus. If that's not fast enough, you might duplicate the dataoe flip-flop and floorplan the copies. Another way would be to add a gclk buffer, floorplan it the buffer that gives the best drive to the data bus pins and floorplan the dataoe flip-flop to be in that corner of the die. -- Phil Hays "Irritatingly, science claims to set limits on what we can do, even in principle." Carl SaganArticle: 18561
martin@the-thompsons.freeserve.co.uk wrote: > On Sat, 30 Oct 1999 16:27:50 -0400, Ray Andraka <randraka@ids.net> > wrote: > > >At only 4 MHz, you are leaving alot of the FPGA's potential performance on > >the table. If your sample rate only needs 4 MHz, you ought to consider bit > >serial design so that you can get into a smaller and cheaper part, for > >example the Xilinx spartan series. Bit serial hardware is typically about > >1/nth the hardware for an n bit wide process, although it also requires an n > >times bit clock. you can use a hybrid digit-serial approach if the clock is > >too high. > > > > My point was that smaller/cheaper is not the issue for me at the > moment, once smaller/cheaper becomes important, we'll be in ASIC > territory anyway. At the prices of the spartan devices, I think you may be pleasantly surprised as to the cost compared with an ASIC. At under $10 in quantity, it requires a fairly large production run with an ASIC to break even. Nice thing with an FPGA solution is that is a lot easier and cheaper to make changes later if you need/want to. > > > >Altera does tend to do better with uninspired synthesis. That is mostly a > >consequence of the global routing structure which makes placement > >considerably less critical than xilinx. However, that is a double edged > >sword: that same global routing limits how fast you can make a design go if > >you are willing to put in the extra effort. > > Presumably also a double edged sword in that if you can't/don't want > to put in the extra effort it makes life harder in 'Xilinx-land' As always. It depends. > <snip lots of useful comparison of architectures> > > > >So, if your application uses delay queues or arithmetic with more than two > >inputs, you are going to use up alot of your logic in ALtera, where you > >wouldn't in xilinx. Thus, you need a device with more "marketing gates" to > >fulfill the same function. Also, the inability to use the LUTs as RAM is a > >handicap if you are doing adaptive filtering, as it keeps you out of some > >significant algorithmic optizations. > > > > I haven't looked hard enough at the architectures to work it out, but > are they not both LUT based? What extras do the Xilinx arch. have > that enables all these other features? And why aren't Altera > interested in copying them :) They are both LUT based, but the xilinx devices allow you to use the LUT as a 16x1 RAM (ie you can change the LUT contents from within the FPGA). That feature, which xilinx holds a patent on, allows you to use the LUT as a 16 bit shift register or as re-writable lookups. There is nothing equivalent in Altera. Most DSP applications have filtering or something similar going on, which in turn implies the fairly extensive use of delay queues. The xilinx LUT ram capability gives you a compact way of doing the delays, where Altera uses up an LE for each clock delay on each bit. For arithmetic circuits, the altera breaks the 4 input LUT into a pair of 3 input LUTs, one for the carry function and one for the 'sum' function. That means if your arithmetic function is more than 2 inputs (an adder-subtractor or an accumulator with clear for example), the altera design is forced to a second level of logic (slower and twice the area). The xilinx carry chain uses additional dedicated logic, which in the case of the 4K/spartan architecture is in front of the 4 input LUT. That leaves you with a 4 input function at each bit whether or not the carry is used. Thus you can realize a 3 input arithmetic function (like ALtera, one input is used for the carry in) in one level of logic. Xilinx Virtex puts the dedicated carry logic after the LUT, so the carry in does not feed the LUT. In that case you can have arithmetic functions with four inputs. Additionally, if you need to use a clock enable, the Altera architecture (10K, not Apex), steals one of the LUT inputs for the clock enable, which reduces the effective size of a LUT to 3 input or in the case of an arithmetic function a one input arithmetic function (let's see - that would be a 2's complement or add a fixed number and not much else). The Xilinx clock enable does not steal a LUT input, so there is no logic penalty for using it. An efficient method for filtering is distributed arithmetic, which essentially puts all the possible 1xn sums for a 4 tap filter into a look-up table. That way you get a serial 4 tap filter in n LUTs (n is the width of your coefficient plus 2 bits). You can combine them both for longer filters and for processing more than one bit of the input per clock. See the latest paper on my website (FPGAs make radar signal processor on a chip a reality) for more detail about how that is done. This same method could be used in adaptive filters if you could reprogram the LUTs to change the coefficients on demand. The xilinx architecture gives you a method to do this since you can use the LUTs as RAM. > >For low data rate stuff, I would prefer to use a smaller device and go with a > >bit serial or digit serial design to keep unit costs down. The exception is > >where the higher clock rates might cause you problems. The extra design > >effort required for that type of architectural optimization is pretty small. > > > > Again, units costs aren't really the issue, although if I can get an > order of magnitude difference in price ... :) > I think you're close to an order of magnitude in required area. The pricing at the lower end is a little compressed, but I think you'll still see a half a magnitude price reduction. Again, we're not talking an absolutely tailored design here, just a change from a bit-parallel to a bit serial solution. > I have to say that I am torn, in that we already use Altera devices > (7K's) for a number of applications, but everything I read here > implies that we need to look at Xilinx for these new applications. > Most of our applications so far have been glue-logic type things > rather than arithmetic based work. Ownership and familiarity of a particular vendor's tool set is a compelling reason to stay with that vendor. Unless you find yourself hitting the wall because of architectural differences, I'd probably stick with the Altera if I were in your shoes (without knowing much about your design). On the other hand, you should at least be aware of what can be done in other architectures so that you have an idea of how much the altera solution is costing you in your system context. Also, it might be worth comparing the cost of ownership of the Xilinx Spartan chips with an Asic for your application. > > > My other problem is that I'm not au fait enough with either the Altera > 10K's or the Xilinx devices to really know how far I could push my > example design if I put the effort in (without actually doing it, > which will involve a significant investment in terms of time and > needing to buy Xilinx tools we may not need). > > Anyway, thanks for the input, > Martin > > >martin@the-thompsons.freeserve.co.uk wrote: > > > >> If I can put a different point of view. > >> > >> I'm new to the game, but I just implemented a signal processing type > >> design, which I wrote as 'behavioural-ish' (ie not aimed directly at > >> being synthesisable) VHDL. It isn't huge, but it has a collection of > >> multipliers/adders/comparators. I thought I'd just see how far off a > >> usefully synthesizable design it was so I ran it through FPGA Express. > >> > >> It runs at 4MHz in a Virtex, and 14MHz in Flex10K (and 7MHz in a > >> Flex6016 - just to see). 4MHz is just fast enough, but the Altera > >> device gives me plenty of spare time to play with. The big plus for > >> me is that I can get the performance I want without having to mess > >> around hand-tuning the algorithm to the architecture, and at the > >> moment, all I'm (or my boss is :-) looking for it proof of concept. > >> > >> It appears from what I've gathered so far - and as I say I am new to > >> this - that Xilinx will perform better, given the up front investment. > >> However, the Altera devices may well perform *well enough* straight > >> out of the box with very 'ordinary' VHDL. > >> > >> So there's another view: The choice may depend on how much > >> time/inclination you have to tweak your implementation. > >> > >> Which brings me to a question... This design fits in a 10K20, which > >> is nominally a 20K gates device, so clearly its not a huge design. Am > >> I going to find the Altera architecture limits me, and will also > >> require hand-tuning, as the design size goes up? > >> > > > > > > > >-- > >-Ray Andraka, P.E. > >President, the Andraka Consulting Group, Inc. > >401/884-7930 Fax 401/884-7950 > >email randraka@ids.net > >http://users.ids.net/~randraka > > > > > > Martin Thompson > martin@the-thompsons.freeserve.co.uk > http://www.the-thompsons.freeserve.co.uk/ -- -Ray Andraka, P.E. President, the Andraka Consulting Group, Inc. 401/884-7930 Fax 401/884-7950 email randraka@ids.net http://users.ids.net/~randrakaArticle: 18562
To be honest, I find the whole constraint thing to be very, very confusing. I think that the way they are set up and more so the way they are documented makes them very, very hard to understand. I find this doubly painful when it seems that the concept of constraints is so simple. But now that my beef is out of the way I will say that I have dealt with this sort of thing using the OFFSET command. In my last Xilinx design (which is starting to grow whiskers) I had a clock, LCLK and a bidir bus LAD. I spec'ed using the statements below to assure that the outputs were stable 5 nS before the rising edge of LCLK and could tolerate changes on LAD up to 20 nS after LCLK. I did it this way because it seemed easier to relate these numbers to the devices I was connecting to (setup = 5 nS and output delay = 20 nS). NET LAD<* OFFSET = IN 20 AFTER LCLK; NET LAD<* OFFSET = OUT 5 BEFORE LCLK; To the best of my knowledge this also took into account the path through the enable on the tristate buffers. I did not need to place any special constraints on that signal. The data book does not show OFFSET being used with IN and AFTER or OUT and BEFORE. Rather they show them used in the opposite combination. But this did what I wanted and seemed to work. Andy Peters wrote: > > I'm using an XC4KE part, I have a 32-bit bidirectional bus that comes out to > the pins and I'd like to constrain the tristate enable so that it's faster. > > There's a constraint called TPSYNC, which "allows defintion of synchronous > points that are <i>not</i> FFS, RAMS, PADS or LATCHES" and they are > "commonly used with three-state buffers." > > The example given in the Xilinx "timing and constraints" presentation is a > bit weak. I have a signal called dataoe, which is a flip-flop output, and > that signal is used as the tri-state enable for the outputs, which are > called ramdata[31:0]. Any hints as to how to write this constraint? Assume > that I want to constrain it to be 10 ns. > > thanx, > > a > > -- > ----------------------------------------- > Andy Peters > Sr Electrical Engineer > National Optical Astronomy Observatories > 950 N Cherry Ave > Tucson, AZ 85719 > apeters (at) noao \dot\ edu > > "Creation Science" is oxymoronic. -- Rick Collins rick.collins@XYarius.com remove the XY to email me. Arius - A Signal Processing Solutions Company Specializing in DSP and FPGA design Arius 4 King Ave Frederick, MD 21701-3110 301-682-7772 Voice 301-682-7666 FAX Internet URL http://www.arius.comArticle: 18563
[snip bit-serial suggestions] > My point was that smaller/cheaper is not the issue for me at the > moment, once smaller/cheaper becomes important, we'll be in ASIC > territory anyway. Won't the same techniques help you get a smaller/cheaper ASIC? -- These are my opinions, not necessarily my employers.Article: 18564
An update is now available for download at www.symphonyeda.com that fixes this problem. The fix is contained in Build#11 Thanx Haneef Allan Herriman <allan.herriman.hates.spam@fujitsu.com.au> wrote in message news:3816b2cb.1582367@newshost.fujitsu.com.au... > On Sat, 23 Oct 1999 10:19:41 -0700, "Haneef D. Mohammed" > <haneef@mindspring.com> wrote: > > > I tried it on an old project (which I had previously compiled with a > few different tools), and it gave an internal error on the very first > file I tried to compile. > > >vhdlp ..\src\ex_1164\ex_1164.vhd > Symphony EDA (R) VHDL Compiler/Simulator Module VhdlP, Version 1.4, > Build#10. > Copyright(C) Symphony EDA 1997-1999. All rights reserved. > Reading C:\Program Files\Symphony EDA\VHDL Simili\bin\symphony.ini ... > Library 'ieee' ==> $SYMPHONY/Lib/Ieee/Ieee.sym > Library 'work' ==> work.sym > Reading $SYMPHONY\Lib\Ieee\Ieee.sym\std_logic_1164\std_logic_1164.var > Parsing Package:exemplar_1164 @ line ..\src\ex_1164\ex_1164.vhd:31 > Writing work.sym\exemplar_1164\exemplar_1164.var > Parsing Package Body:exemplar_1164 @ line > .\src\ex_1164\ex_1164.vhd:297 > Internal Error: D:\home\proj\SimVHDL\VhdlExpr\CSTypeConversion.cpp: > (line 79): E > xpecting Aggregate here > > And the source around line 297 was: > > 290: signal q : out std_ulogic_vector); > 291: > 292:end exemplar_1164; > 293: > 294:library ieee ; > 295:use ieee.std_logic_1164.all ; > 296: > 297:package body exemplar_1164 is > 298: > 299: --- > > > Any idea what is wrong? > > Regards, > Allan.Article: 18565
The Inkjet Company has the following inkjet specials: All Prices Include Free Shipping and Free Sales Tax EPSON Epson - $8.00 - Black or 10 for $65.00 - All Models* Epson - $12.00 - Color or 10 for $100.00 - All Models* *Epson - $16.00 - Black or 10 for $135.00 - 900/3000 *Epson - $19.00 - Color or 10 for $165.00 - 750/900/1200/3000 CANON Canon BCI-21B - $5.00 - Black BJC-2500/4000/4400/5000 Canon BCI-21C - $8.00 - Color BJC-2500/4000/4400/5000 Canon BC-20 - $18.00 - Black BJC-2500/4000/4300/4400/ Canon BC-02 - $16.50 - Black BJC-210/230/BJ-200 Canon BC-05 - $28.00 - Tri-Color BJC-210/240 Canon BX-3 - $28.00 - Black B-540/550/640 (See Below For Apple & H.P. Products) The Inkjet Company 1-310-517-9999 1-503-905-6652 Fax APPLE M8041G/C - $17.00 - Black Styler Writer 1500 M4609G/C - $30.00 - Tri-Color Styler Writer 1500 M3240G/A - $19.00 - Large "Double" Black Style Writer 2400 M3330G/A - $5.00 - Black Style Writer 2400 M3329G/A - $9.00 - Tri-Color Style Writer 2400 M1949G/A - $3.75 - Cyan Color Style Writer Pro M1950G/A - $3.75 - Magenta Color Style Writer Pro M1951G/A - $3.75 - Yellow Color Style Writer Pro M1952G/A - $3.75 - Black High Capcity Color Style Writer Pro HEWLETT PACKARD 51625A - $19.00 - Tri-Color Deskjet 500c/550 Series 51626A - $16.50 - Black Deskjet Series / OfficeJet 300 / HP Fax 51629A - $16.50 - Black Deskjet 600 Series / OfficeJet 500 Series 51649A - $19.00 - Tri-Color Deskjet 600 Series / OfficeJet 500 Series 51645A - $16.50 - Black Deskjet 700/800/1000/1600 51640A - $16.50 - Black Deskjet 1200 / CopyJet The Inkjet Company 1-310-517-9999 1-503-905-6652 Fax Visa, Mastercard, Amex, Discover Fast Free Delivery - usually in two days or less!Article: 18566
Hi, Using a Flex10K and have some code which has a bunch of FF's that reset, and a few that preset. Now I understand that the Flex will power up it's FF's in a low state, and it's up to a reset input to perform a functional reset once the device has configured. I also have a processor connected to the device and want to hold it in reset until the FPGA has configured. I have looked at using either INIT_DONE or CONF_DONE, but one de-actives before the end of config, and the other is inactive for a period at the start of config. I could also set a pin on the FPGA permanently high in the verilog, and us a pull-down externally so that during configuration it goes low, and when the user i/o is enabled, it gets driven high ... this could be my processor reset. I then connect a separate reset line from the processor to the fpga as the functional reset .... ... sounds abit clunky though ... is there a more elegant way of doing this? Thanks, Gary Cook, Sony Oxford, UK.Article: 18567
I have a problem with Lattice ispExpert 7.1.
I wrote this very simple 16 bit counter in ABEL-HDL
MODULE counter16
count0..count15 pin istype 'reg';
countck pin;
EQUATIONS
[count0..count15].clk = countck;
[count0..count15] := [count0..count15] + 1;
END
Someone can explain me why it requires 18 macrocells instead 16?
I have to set some option in the compiler?
Besides, if I write
[count0..count15] := [count0..count15] + extpin;
where extpin is a input pin assuming 0 or 1 values, this requires
only 17 macrocells!!!
Thanks in advance.
Luigi
Article: 18568The Inkjet Company has the following inkjet specials: All Prices Include Free Shipping and Free Sales Tax EPSON Epson - $8.00 - Black or 10 for $65.00 - All Models* Epson - $12.00 - Color or 10 for $100.00 - All Models* *Epson - $16.00 - Black or 10 for $135.00 - 900/3000 *Epson - $19.00 - Color or 10 for $165.00 - 750/900/1200/3000 CANON Canon BCI-21B - $5.00 - Black BJC-2500/4000/4400/5000 Canon BCI-21C - $8.00 - Color BJC-2500/4000/4400/5000 Canon BC-20 - $18.00 - Black BJC-2500/4000/4300/4400/ Canon BC-02 - $16.50 - Black BJC-210/230/BJ-200 Canon BC-05 - $28.00 - Tri-Color BJC-210/240 Canon BX-3 - $28.00 - Black B-540/550/640 (See Below For Apple & H.P. Products) The Inkjet Company 1-310-517-9999 1-503-905-6652 Fax APPLE M8041G/C - $17.00 - Black Styler Writer 1500 M4609G/C - $30.00 - Tri-Color Styler Writer 1500 M3240G/A - $19.00 - Large "Double" Black Style Writer 2400 M3330G/A - $5.00 - Black Style Writer 2400 M3329G/A - $9.00 - Tri-Color Style Writer 2400 M1949G/A - $3.75 - Cyan Color Style Writer Pro M1950G/A - $3.75 - Magenta Color Style Writer Pro M1951G/A - $3.75 - Yellow Color Style Writer Pro M1952G/A - $3.75 - Black High Capcity Color Style Writer Pro HEWLETT PACKARD 51625A - $19.00 - Tri-Color Deskjet 500c/550 Series 51626A - $16.50 - Black Deskjet Series / OfficeJet 300 / HP Fax 51629A - $16.50 - Black Deskjet 600 Series / OfficeJet 500 Series 51649A - $19.00 - Tri-Color Deskjet 600 Series / OfficeJet 500 Series 51645A - $16.50 - Black Deskjet 700/800/1000/1600 51640A - $16.50 - Black Deskjet 1200 / CopyJet The Inkjet Company 1-310-517-9999 1-503-905-6652 Fax Visa, Mastercard, Amex, Discover Fast Free Delivery - usually in two days or less!Article: 18569
Can anybody tell me how to do capture and readback though a Virtex's JTAG port ? As far as I can see this is possible but is not supported by either the Xilinx JTAG programmer or the H/W debugger. I don't have access to the SelectMAP port so I probably need some s/w to drive the JTAG port via e.g a parallel cable. I'm also very confused as to how to define and get at the user definable internal scan chains. One possible route seems to be through the JAVA based B-scan API. Anybody tried this ?Article: 18570
Hi Andy, try to write in the your ucf file something like this: NET ramdata(* TPSYNC=TRI_STATE_OUTPUTS; TIMESPEC TS_OE_signal=FROM:FFS(dataoe):TO:TRI_STATE_OUTPUTS=10ns; It will constrain path time from OE signal driver to your tri-state bus outputs. If you want constrain path time from dataoe flip-flop to output enable pin of OBUFE then you shoul attach TPSYNC attribute on these pins directly. Hope it will be small help for you. Andrey Ushenin, -------------------------- DSP laboratory, Novgorod State University, Russia Andy Peters <apeters.nospam@nospam.noao.edu.nospam> wrote in message news:7vdc8m$1uij$1@noao.edu... > I'm using an XC4KE part, I have a 32-bit bidirectional bus that comes out to > the pins and I'd like to constrain the tristate enable so that it's faster. > > There's a constraint called TPSYNC, which "allows defintion of synchronous > points that are <i>not</i> FFS, RAMS, PADS or LATCHES" and they are > "commonly used with three-state buffers." > > The example given in the Xilinx "timing and constraints" presentation is a > bit weak. I have a signal called dataoe, which is a flip-flop output, and > that signal is used as the tri-state enable for the outputs, which are > called ramdata[31:0]. Any hints as to how to write this constraint? Assume > that I want to constrain it to be 10 ns. > > thanx, > > a > > -- > ----------------------------------------- > Andy Peters > Sr Electrical Engineer > National Optical Astronomy Observatories > 950 N Cherry Ave > Tucson, AZ 85719 > apeters (at) noao \dot\ edu > > "Creation Science" is oxymoronic. > >Article: 18571
On Sun, 31 Oct 1999 20:51:53 -0800, "Haneef D. Mohammed" <haneef@mindspring.com> wrote: >An update is now available for download >at www.symphonyeda.com that fixes this problem. >The fix is contained in Build#11 Thanks, it fixed that bug, but now there's another... BTW, int2evec returns a 32 element slv if the second argument isn't given. (This is based on old code. I wouldn't use this function now.) library ieee; use ieee.std_logic_1164.all; library work; use work.exemplar_1164.all; -- you already have this entity foobar is end foobar; architecture snafu of foobar is signal foo : std_logic_vector(7 downto 0); signal bar : integer; begin foo <= int2evec(bar, 8); -- this line is ok foo <= int2evec(bar)(7 downto 0); -- this line gives an error end snafu; C:\>vhdlp tmp.vhd Symphony EDA (R) VHDL Compiler/Simulator Module VhdlP, Version 1.4, Build#11. Copyright(C) Symphony EDA 1997-1999. All rights reserved. Reading C:\tools\Symphony EDA\VHDL Simili\bin\symphony.ini ... Library 'ieee' ==> $SYMPHONY/Lib/Ieee/Ieee.sym Library 'work' ==> work.sym Reading $SYMPHONY\Lib\Ieee\Ieee.sym\std_logic_1164\std_logic_1164.var Reading work.sym\exemplar_1164\exemplar_1164.var Parsing Entity:foobar @ line tmp.vhd:8 Writing work.sym\foobar\foobar.var Parsing Architecture:foobar(snafu) @ line tmp.vhd:11 Error: CSVHDL0149: tmp.vhd: (line 20): Bad assignment. Expression on RHS is not of type ''std_logic_vector(7 DOWNTO 0)'' Errors: 1, Warnings: 0, Elapsed Time: 00h:00m:00s:736ms C:\> FYI, Allan.Article: 18572
Phil Hays wrote in message <381C922B.3603C633@sprynet.com>... >Andy Peters wrote: > >> I'm using an XC4KE part, I have a 32-bit bidirectional bus that comes out to >> the pins and I'd like to constrain the tristate enable so that it's faster. >> ... I have a signal called dataoe, which is a flip-flop output, and >> that signal is used as the tri-state enable for the outputs, which are >> called ramdata[31:0]. > >I have usually not had very good luck getting the Xilinx tools to improve timing >in a case like this by adding timing constraints. What I would suggest is that >you floorplan the dataoe flip-flop to the middle of the tristate bus. If that's >not fast enough, you might duplicate the dataoe flip-flop and floorplan the >copies. Another way would be to add a gclk buffer, floorplan it the buffer that >gives the best drive to the data bus pins and floorplan the dataoe flip-flop to >be in that corner of the die. Unfortunately, the pins are fixed, but I'll try floorplanning the dataoe flop as you say. I added a BUFGS to dataoe, and for some reason, it got slower! i think I screwed something up, though, since it was 6:30 pm on a Friday... -- a ----------------------------------------- Andy Peters Sr Electrical Engineer National Optical Astronomy Observatories 950 N Cherry Ave Tucson, AZ 85719 apeters (at) noao \dot\ edu "Creation Science" is oxymoronic.Article: 18573
Where can I find information about the roots of development of FPGA's and evolution untill nowadays? Diferences between others systems? .... * Sent from AltaVista http://www.altavista.com Where you can also find related Web Pages, Images, Audios, Videos, News, and Shopping. Smart is BeautifulArticle: 18574
Rickman wrote in message <381CB6F4.1F6908AF@yahoo.com>... >To be honest, I find the whole constraint thing to be very, very >confusing. I think that the way they are set up and more so the way they >are documented makes them very, very hard to understand. I find this >doubly painful when it seems that the concept of constraints is so >simple. and the constraints editor GUI, which is supposed to make life easier, still has problems: it dies with an access fault every time you exit, it screws up the text file so you any kind of neat ordering you'd created gets blown away, and it doesn't have a way to add all the possible constraints. I've stopped using it. >But now that my beef is out of the way I will say that I have dealt with >this sort of thing using the OFFSET command. In my last Xilinx design >(which is starting to grow whiskers) I had a clock, LCLK and a bidir bus >LAD. I spec'ed using the statements below to assure that the outputs >were stable 5 nS before the rising edge of LCLK and could tolerate >changes on LAD up to 20 nS after LCLK. I did it this way because it >seemed easier to relate these numbers to the devices I was connecting to >(setup = 5 nS and output delay = 20 nS). > >NET LAD<* OFFSET = IN 20 AFTER LCLK; >NET LAD<* OFFSET = OUT 5 BEFORE LCLK; > >To the best of my knowledge this also took into account the path through >the enable on the tristate buffers. I did not need to place any special >constraints on that signal. > >The data book does not show OFFSET being used with IN and AFTER or OUT >and BEFORE. Rather they show them used in the opposite combination. But >this did what I wanted and seemed to work. OFFSET IN assumes that the FPGA is getting data from an upstream device. OFFSET IN BEFORE specifies that the data will be valid at the FPGA input pins x ns before the clock arrives AT THE FPGA. OFFSET IN AFTER specifies that the data will be valid at the FPGA input pins x ns AFTER the clock arrives at the UPSTREAM DEVICE. OFFSET IN AFTER is probably more useful. If your FPGA is driven by an octal flipflop with a clock-to-out time of 5 ns, all you need to do is specify NET din OFFSET = IN 5 ns AFTER clock; and the tools should (hopefully) take care of the rest. OFFSET OUT is similar, 'cept that the FPGA is the data source, rather than the sync. OFFSET OUT AFTER specifies that the data will be valid at the FPGA OUTPUT x ns AFTER the clock arrives at the FPGA INPUT. OFFSET OUT BEFORE specifies that the data will be valid AT THE FPGA OUTPUT x ns BEFORE the clock arrives at the DATA SINK (upstream device). OFFSET OUT BEFORE would be used to specify a setup time for the data sink (you need to include any board delay). OFFSET OUT AFTER would be used to constrain the FPGA outputs to get the data out in a certain time, but you have to make sure that you're still meeting the data sink's setup requirements. hope this helps... -- a ----------------------------------------- Andy Peters Sr Electrical Engineer National Optical Astronomy Observatories 950 N Cherry Ave Tucson, AZ 85719 apeters (at) noao \dot\ edu "Creation Science" is oxymoronic.
Site Home Archive Home FAQ Home How to search the Archive How to Navigate the Archive
Compare FPGA features and resources
Threads starting:
Authors:A B C D E F G H I J K L M N O P Q R S T U V W X Y Z