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
> I also upgraded to Webpack 6.2 and got quite a shock. For the sp3, the > fmax has shot up to 430MHz overall and the 2 larger blocks besides > blockram are in the 550MHz ballpark. If true, then Wow! Your FPGA based CPU is faster than 0.13 ASIC CPUs. The timing report is essentially > same as before, no special warnings but some improvement in compile > time I was looking for. Is that post-synthesis or post-layout? > As a crosscheck, I redid the other sp2, v2pro > and got same results as before. Only the sp3 goes up. 430MHz x 1.3 is > about 550mips. On my own designs, I only saw a small increase in sp3 performance when upgrading to 6.2. But then I don't use XST, so maybe that's whats giving you the increase. Have you tried other synthesis tools / physical optimisation tools? They should give you even more performance. If you really want to verify it, why don't you run a back annotated gate-level sim? > I have been contemplating how to license the design, I would like to > see the final design in its Transputer form open to individuals and > dot edu, but $ for commercial interest. Trolltech does this with Qt, > gpl for non commercial Linux and the KDE world while still doing well > for $ Windows market. > > I am interested in other points of view on this. Sounds good. One day i intend to license my own (more humble) CPU in this way. Make sure to use GPL rather than LPGL. Cheers, JonArticle: 67401
sounds reasonable... Kelvin "Kevin Neilson" <kevin_neilson@removethiscomcast.net> wrote in message news:xEQ3c.1551$zS4.16753@attbi_s51... > The design is 100% routed but still doesn't meet timing until the number in > parentheses is zero. The router has to keep ripping up and rerouting paths > until timing is met. > -Kevin > > "Kelvin @ SG" <kelvin8157@hotmail.com> wrote in message > news:404fbb0f$1@news.starhub.net.sg... > > Hi, there: > > > > I have been confused why ISE6 always spend a lot of time after it has > > completed routing... > > On the routing log, it seems at Phase 7, it accomplished all routing > > already...but why Phase 8? > > > > Best Regards, > > Kelvin > > > > > > > > Phase 9.27 > > Phase 9.27 (Checksum:55d4a77) REAL time: 5 mins 4 secs > > > > Writing design to file switch_top.ncd. > > > > Total REAL time to Placer completion: 5 mins 7 secs > > Total CPU time to Placer completion: 5 mins 3 secs > > > > > > Phase 1: 13137 unrouted; REAL time: 5 mins 10 secs > > > > Phase 2: 12564 unrouted; REAL time: 5 mins 57 secs > > > > Phase 3: 3148 unrouted; REAL time: 6 mins 3 secs > > > > Phase 4: 3148 unrouted; (22384) REAL time: 6 mins 4 secs > > > > Phase 5: 3161 unrouted; (22208) REAL time: 6 mins 5 secs > > > > Phase 6: 3161 unrouted; (22208) REAL time: 6 mins 6 secs > > > > Phase 7: 0 unrouted; (22732) REAL time: 6 mins 27 secs > > > > Writing design to file switch_top.ncd. > > > > Phase 8: 0 unrouted; (22195) REAL time: 10 mins 24 secs > > > > > > > >Article: 67402
Dear colleagues, I'm interested if any of you have had any practical experience with implementing gigabit protocols with Altera Startix GXs. Any information on good or bad experience would be truly appreciated. Also, did you use Altera's GX-evaluation board ever. How did you find it ? Thanks in advance, VladArticle: 67403
On Thu, 11 Mar 2004 16:35:22 +0800, Kelvin @ SG wrote: >what is the oftenly used algorithms for RC4 encryption in an FPGA? Same as in software. The algorithm for RC4 is published by RSA Labs, but how you implement it is up to you. >is it true that i have to use at least 256X8X2 registers or RAM? It's a permutation cypher. I can't remember the block size offhand, but that sounds about right (particularly with block-chaining). It's quite a modest RAM requirement, as crypto algorithms go. -- MaxArticle: 67404
Kevin Neilson <kevin_neilson@removethiscomcast.net> wrote: : I know; this is off-topic: I can't figure out why, when there is a brief : power outage, my answering machine loses time information but retains voice : messages. I guess the messages must be stored in FLASH. Is it just too : much of a pain to write the time into FLASH as well? Only you know it was a "brief" power outage. Arguably storing time information would be a bad thing, because after a long outage the machine would come up with a 'valid' but completely wrong time. Another possible reason is the undesirability of performing regular writes to FLASH memory (e.g. every second), either because of the potential for 'fatigue' failure of the device or because of risks associated with the power failing in the middle of a write or erase operation. Richard. http://www.rtrussell.co.uk/ To reply by email change 'news' to 'richard'Article: 67405
I am trying this: module uv_v_filter ( input clk, input active, input[7:0] y, input[7:0] u, input[7:0] v, output activeout, output[7:0] yout, output[7:0] uout, output[7:0] vout ); parameter log2samplesline=9; reg[log2samplesline-1:0] samplecounter; reg[7:0] yout, uout, vout; integer i; // line memories reg activeout; reg[7:0] yshift0[0:(1<<log2samplesline)-1]; reg[7:0] ushift0[0:(1<<log2samplesline)-1]; reg[7:0] ushift1[0:(1<<log2samplesline)-1]; reg[7:0] vshift0[0:(1<<log2samplesline)-1]; reg[7:0] vshift1[0:(1<<log2samplesline)-1]; always @(posedge clk) begin yout<=yshift0[samplecounter]; uout<=( u * 1 + ushift0[samplecounter] * 2 + ushift1[samplecounter] * 1 + 2 )>>2; vout<=( v * 1 + vshift0[samplecounter] * 2 + vshift1[samplecounter] * 1 + 2 )>>2; // shift registers / line memories if (active) begin ushift0[samplecounter]<=ushift1[samplecounter]; vshift0[samplecounter]<=vshift1[samplecounter]; yshift0[samplecounter]<=y; ushift1[samplecounter]<=u; vshift1[samplecounter]<=v; samplecounter<=samplecounter+1; end else begin samplecounter<=0; end activeout<=active; end endmodule And Quartus seems to optimize the memories away to a strange number of 22 bit's, targeting a cyclone. And funny is: vshift1 gets 20 bits. yshift0 gets 2 bits Is Quartus or me the problem ? Raymund HofmannArticle: 67406
I doubt that this might be the problem, in the START state I make a first check to make sure this problem does not occur. I don't think it's possible for count to be less than 1.Article: 67407
> What exactly is the purpose of this CPU, that is, why would a user want > this CPU instead of a commercial CPU? It sounds like it uses a fair > amount of FPGA real estate, although the speed is impressive. But it > requires a memory interface that can keep it fed which will use a lot of > IOs. > > Also, what do you mean about it's "Transputer form"? > > -- > > Rick "rickman" Collins > > Hi Rickman The original and still main intent of the project is to create a new Transputer. The underlying cpu is just a std RISC with hyperthreading but the Transputer specifics are process creation and message support in HW to allow par type programming at very fine granularity as well as scalability to allow unlimited cpu ganging. The ganging is based on the simple idea that Transputers are just widgets like any other chips to be glued together into larger cpus. It works because of the message passing scheme being similar to HW. The 2 parts can be separated, ie Transputer like features can be added to any cpu with KROC port of Occam, but most cpus have very high swapping context overhead to make it work as well and are missing alot of the plug an play HW, like links. These specifics can be added onto the base cpu when it is stable. I may release them as separate cpus. You are right about the memory bandwidth. I am planning on using RLDRAM, a few x $ of DDR, but way more bandwidth, and way less latency. IE 20ns cycles at 8 banks. 1 bank per cpu seems right as 2nd level cache. DDR for primary if more needed. The real idea behind Transputers is to distribute cpus into the memory space so 24b address space may be fine. Adding 1G memory to each Transputer is like saying I don't want more computing. Before FPGAs took of, Transputers were the way to do alot of the same things (at the higher end) since Occam (Inmos transputer language) was sort of a HDL for processes. Occam has since evolved into HandelC, same people but using FPGAs directly. The technology died out because Inmos didn't have their act together and FPGAs weren't around to help them prototype. The 2 tchnologies crossed paths without barely knowing each other. Most all the older European FPGA houses today came from Transputer backgrounds ie Nallatech, Sundance & many others. Many of the earlier super computers used i860, Sparcs, Alphas or whatever as their compute nodes, but used Transputers for their communications into the backplane. Any Transputer person today seeing PCI Express, Hypertransport, hyperthreading, and a dozen other link like technologies would smile on, already did that 20yrs ago. Why am I doing this, well I am interested in bringing the Transputer technology back from the dead, and FPGAs esp sp3 looks incredible to just waste on conventional cpu design. I also worked at Inmos so I have the perspective of having beeing part of something great that made some mistakes. I'd give it a 2nd chance in a heartbeat. Not that many people around would have the confidence to do this, but I wear many hats, just takes time. I want my own Bluegene, I can't afford one. So I put together N cpus to do the kinds of things only Transputer farms can do, buts thats another future project. If I scale N Transputers together, the cost is linear to the basic HW costs. Perhaps sp3-5000 can hold 20+ instances, but theres power issues to deal with. Ofcourse if I got to 250K units, I might get a price break too. Now when was the last time anyone offered an Athlon in N way cpu that scales by instance count. It doesn't work for most cpus since they are stuck with the shared data model and expensive NRE for special smp controllers. A couple of slower Transputers should easily be able to beat 1 faster x86 at any task. Even Intel is seeing the light. But Intel doesn't understand Occam and is stuck with x86 rock around its neck. regards all johnjakson_usa_com sorry to sound like blowing my own horn, but the possibilities of FPGA technology seems far more interesting than the endless x86 drivel on comp.arch.Article: 67408
Bruno a écrit: > I doubt that this might be the problem, in the START state I make a > first check to make sure this problem does not occur. I don't think it's > possible for count to be less than 1. I'm afraid XST (and any other tool) is not clever enough to know this. Besides, if length = 1, count will be 0. Try it :o) -- ____ _ __ ___ | _ \_)/ _|/ _ \ Adresse de retour invalide: retirez le - | | | | | (_| |_| | Invalid return address: remove the - |_| |_|_|\__|\___/Article: 67409
"Kevin Neilson" <kevin_neilson@removethiscomcast.net> wrote in message news:BJQ3c.1317$C51.24342@attbi_s52... > I know; this is off-topic: I can't figure out why, when there is a brief > power outage, my answering machine loses time information but retains voice > messages. I guess the messages must be stored in FLASH. Is it just too > much of a pain to write the time into FLASH as well? If it can't be written > into FLASH, couldn't it use a small cap to back up the volatile RAM? Maybe > I shouldn't expect too much for $25. > -Kevin > I find this particular issue a PITA too, especially as I have to find the damn user manual everytime I want to set the time on my answerphone - #*55 or whatever I have to press to get it into time-set mode. Rather than FLASH, wouldn't the normal approach be to fit battery-backup for the real-time clock? It would mean very occassional battery replacement but it would probably last 5 years or so, by which time they'd likely call the product obsolete ;) Jim -----= Posted via Newsfeeds.Com, Uncensored Usenet News =----- http://www.newsfeeds.com - The #1 Newsgroup Service in the World! -----== Over 100,000 Newsgroups - 19 Different Servers! =-----Article: 67410
rickman <spamgoeshere4@yahoo.com> wrote in message news:<404FE5F4.146E6CE4@yahoo.com>... > john jakson wrote: > > > > A while back I mentioned I'd post an update on progress for this R3 > > hyperthreaded cpu project. > > > ...... > > > > Licensing > > I have been contemplating how to license the design, I would like to > > see the final design in its Transputer form open to individuals and > > dot edu, but $ for commercial interest. Trolltech does this with Qt, > > gpl for non commercial Linux and the KDE world while still doing well > > for $ Windows market. > > > > I am interested in other points of view on this. > I forgot to say the following. If sp3 really can do what Webpack 6.2 says it can, I think we have just crossed another of those unmarked milestones in history where FPGAs can actually do more than an ASIC team can inspite of the fact that FPGAs only simulate HW fabric. It is now possible for cpu designers to build computers with far more interesting architectures in far less time than ASIC teams could ever concieve. FPGA foundry does all the hard work of making the BlockRams and logic fabric go as fast as most ASIC cell designs and architects put it together. Just how fast does nVidia gpu work, only at 300MHz or so, but they probably get to use a couple of times more logic per pipeline, but I bet their memories are still 300MHz. There aren't that many ASIC guys working on anything faster than that. The only thing ASICs still have going is the sheer possibility to put even vaster amounts of stuff onto chips and at lower costs if marketing can sell millions of em. congrats to sp3 if its true.Article: 67411
johnjakson@yahoo.com (john jakson) writes: > A couple of slower Transputers should easily be able to beat 1 > faster x86 at any task. Hmmm, not at _any_ task, I'd say. Some tasks might not be parallelizable enough. And even if they are, you still need people to write these parallel programs, which is very much harder than writing sequential programs. I don't see anyone rewriting their HDK synthesis tools, for example, to take advantage of massive fine-grain parallelism to speed them up. Heck, they don't even take advantage of the widely existing dual-CPU SMP, right? But, your project sounds mighty cool. By all means, do it, and release it with the GPL! :-)Article: 67412
jon@beniston.com (Jon Beniston) wrote in message news:<e87b9ce8.0403110225.313772a3@posting.google.com>... > > I also upgraded to Webpack 6.2 and got quite a shock. For the sp3, the > > fmax has shot up to 430MHz overall and the 2 larger blocks besides > > blockram are in the 550MHz ballpark. > > If true, then Wow! Your FPGA based CPU is faster than 0.13 ASIC CPUs. > Only because of HT and 3 level of LUT logic. HT means the pipelines are mostly independant but cyclic like the spokes of a wheel. Just like DSP (which is what I have been doing alot of over the years since my Inmos days). > The timing report is essentially > > same as before, no special warnings but some improvement in compile > > time I was looking for. > > Is that post-synthesis or post-layout? > Post synth. > > As a crosscheck, I redid the other sp2, v2pro > > and got same results as before. Only the sp3 goes up. 430MHz x 1.3 is > > about 550mips. > > On my own designs, I only saw a small increase in sp3 performance when > upgrading to 6.2. But then I don't use XST, so maybe that's whats > giving you the increase. Have you tried other synthesis tools / > physical optimisation tools? They should give you even more > performance. > Perf is surely due to 3 levels of LUT or 2 mux levels etc. Don't have Symplify or full ISE budget, but would gladly accept a free license if anyone is listening. email at bottom. > If you really want to verify it, why don't you run a back annotated > gate-level sim? > Will do as soon as Verilog source is complete. The Visual C env is much easier to set up testing for partial blocks than any Verilog env I used. C can be used to create "should do same as fn()s" in a jiffy. Can also link it into other tools like the upcoming assembler, compiler, ucf maker etc etc. Also C allows me to create complex logic in C for getting something to work 1st, when it does, turn it back into HDl RTL form. Worked well for me for 20yrs. > > I have been contemplating how to license the design, I would like to > > see the final design in its Transputer form open to individuals and > > dot edu, but $ for commercial interest. Trolltech does this with Qt, > > gpl for non commercial Linux and the KDE world while still doing well > > for $ Windows market. > > > > I am interested in other points of view on this. > > Sounds good. One day i intend to license my own (more humble) CPU in > this way. Make sure to use GPL rather than LPGL. > > Cheers, > Jon Not a big fan of GPL. If no $ licenses are forthcoming, I will fold the them both into MIT/BSD and let the world rip into it. johnjakson_usa_comArticle: 67413
john jakson wrote: > Also C allows me to create complex logic > in C for getting something to work 1st, when it does, turn it back > into HDl RTL form. Worked well for me for 20yrs. Or do it in Pascal/Delphi and you are just a hop and a skip from VHDL. And if _that_ isn't a controversial sentence....Article: 67414
Symon, Other than to set the level to 1X, or 2X, and to trim the current sources, that is all there is. It is not a programmable D/A converter. I am afraid that all your bit twiddling will do you no good at all. Austin Symon wrote: >Austin, >Cool as chips! So, how about letting us know on CAF what those mystery >bits actually do? It seems that they control both the common mode >voltage and the differential voltage. But, which bits control which >characteristics? Please don't make me experiment myself! You could >save me a host of AC coupling caps! > >Of course, I don't expect guaranteed specs. However, I'm surprised >that Xilinx doesn't tout this as a unique selling point, especially as >there are so many differential signalling standards emerging as supply >voltages continue to fall. > >Thanks, Symon. > > >Austin Lesea <austin@xilinx.com> wrote in message news:<c2npfg$dkt1@cliff.xsj.xilinx.com>... > > >>Symon, >> >>Yes, it does, and yes they all work similarly. >> >>Austin >> >> >>Article: 67415
In article <adb3971c.0403110735.71e1b6e8@posting.google.com>, john jakson <johnjakson@yahoo.com> wrote: >jon@beniston.com (Jon Beniston) wrote in message >news:<e87b9ce8.0403110225.313772a3@posting.google.com>... >> > I also upgraded to Webpack 6.2 and got quite a shock. For the sp3, the >> > fmax has shot up to 430MHz overall and the 2 larger blocks besides >> > blockram are in the 550MHz ballpark. >> >> If true, then Wow! Your FPGA based CPU is faster than 0.13 ASIC CPUs. >> > >Only because of HT and 3 level of LUT logic. HT means the pipelines >are mostly independant but cyclic like the spokes of a wheel. Just >like DSP (which is what I have been doing alot of over the years since >my Inmos days). Actually, thats not hyperthreading/SMT, thats interleaved-multithreading or C-slowing. See Chapter 11 and Appendix B: http://www.cs.berkeley.edu/~nweaver/nweaver_thesis.pdf It is a big win in FPGA-based CPUs if you can't get your forwarding path small enough. -- Nicholas C. Weaver nweaver@cs.berkeley.eduArticle: 67416
Fanout/fanin are terms that appear everywhere in the FPGA literatures but I can not really find out what it exactly means. Is someone can enlighten me on this subject ? Thanks. Christophe.Article: 67417
Hi, I have been working for a little time with Altera Cyclone part and I have my project almost ready to go. I am in the process of cleaning things up and I need to review all warnings I got from the building process. One of them is "Warning: Pin XXX not connected" - it is right, I have a track with signal on PCB going into the Altera part but later on I decided to not use it in my FPGA design... What is the common practice of dealing with such pins to clear this warning and do not short this pin with signal to ground/vcc? I hate the practice of ignoring any "stupid" warnings because it quiets my sensitivity for wornings which are potential errors. When I write any piece of software I try to clean it the way no warning is reported during the build process. I want the same happened in Quartus software during my FPGA design synthesis. Thanks for any input.Article: 67418
raymund hofmann wrote: > And Quartus seems to optimize the memories away to a strange number of 22 > bit's, targeting a cyclone. > And funny is: > vshift1 gets 20 bits. > yshift0 gets 2 bits > > Is Quartus or me the problem ? Consider writing a verilog testbench and siming your code. If it doesn't sim as expected, then debug the code until it sims ok. -- Mike TreselerArticle: 67419
> Bruno a écrit: > >> I doubt that this might be the problem, in the START state I make a >> first check to make sure this problem does not occur. I don't think >> it's possible for count to be less than 1. Try running a modelsim compile (vcom test.vhd). This will give you better syntax error messages. -- Mike TreselerArticle: 67420
"chris" <ccoutand@hotmail.com> wrote in message news:33923a80.0403110803.4f889814@posting.google.com... > Fanout/fanin are terms that appear everywhere in the FPGA > literatures but I can not really find out what it exactly means. > Is someone can enlighten me on this subject ? > Thanks. > Christophe. Fanout basically means the number of inputs that may be driven by an output. An excessive number of inputs will load down the output and it won't work reliably. LeonArticle: 67421
rickman <spamgoeshere4@yahoo.com> wrote in message news:<404FDFE5.4961E719@yahoo.com>... > I am aware > that you market an HDL version of an older Forth processor, but I expect > that is not well optimized for FPGA implementation. > From what I've seen on the MPE web site, the RTX2000 core runs reasonably fast in an FPGA. About the same speed as my processor. The stacks use block RAM, so I would expect it to fit reasonably well in an FPGA. Maybe Stephen can tell us how many Spartan II slices it uses. The RTX2000 is an old design, but newer isn't necessarily better. It was designed back when hardware was expensive and it turned out very well. -- BradArticle: 67422
Hi, I have an Avnet V2P7 development board, it is so strange that I can not load my design via System Ace CF card, the "Sys Ace Error" light is always on. The board layout is not complex in that SysAce is the 1st device in JTAG chain, V2P7 as the 2nd, then a PROM and Spart2e. So there should be no trick in generating ACE file in iMPACT: just add the design file, and that's it. Does anyone else seeing similar problem? I am using ISE 6.2 on Windows 2000. Thanks a lot!Article: 67423
> Can the webpack tools be used to synthesize the design to an XNF or EDIF > file to use in the ISE 4.1 tool? No, not for the XC4000 families. That is why there is no synthesis tool in the Classic version of the software. XST can only target Virtex/Spartan-II and later architectures and was never made to work with the older devices. If it was that easy, we would have just thrown the synthesis tool in there. -- BrianArticle: 67424
john jakson wrote: > > rickman <spamgoeshere4@yahoo.com> wrote in message news:<404FE5F4.146E6CE4@yahoo.com>... > > john jakson wrote: > > > > > > A while back I mentioned I'd post an update on progress for this R3 > > > hyperthreaded cpu project. > > > > > ...... > > > > > > Licensing > > > I have been contemplating how to license the design, I would like to > > > see the final design in its Transputer form open to individuals and > > > dot edu, but $ for commercial interest. Trolltech does this with Qt, > > > gpl for non commercial Linux and the KDE world while still doing well > > > for $ Windows market. > > > > > > I am interested in other points of view on this. > > > > I forgot to say the following. > > If sp3 really can do what Webpack 6.2 says it can, I think we have > just crossed another of those unmarked milestones in history where > FPGAs can actually do more than an ASIC team can inspite of the fact > that FPGAs only simulate HW fabric. > > It is now possible for cpu designers to build computers with far more > interesting architectures in far less time than ASIC teams could ever > concieve. FPGA foundry does all the hard work of making the BlockRams > and logic fabric go as fast as most ASIC cell designs and architects > put it together. > > Just how fast does nVidia gpu work, only at 300MHz or so, but they > probably get to use a couple of times more logic per pipeline, but I > bet their memories are still 300MHz. There aren't that many ASIC guys > working on anything faster than that. The only thing ASICs still have > going is the sheer possibility to put even vaster amounts of stuff > onto chips and at lower costs if marketing can sell millions of em. > > congrats to sp3 if its true. I will say your goals are optimistic at least. If you are doing this as a hobby, then fine. But if you are serious about marketing a CPU or core, then you have a long road ahead of you. But then maybe you realize that and are up to the task. One thing I don't quite get about the idea of everyone rolling their own CPU is the effort required. Clearly the automotive companies have a high enough volume to justify designing their own CPUs and SOCs. But mostly they just go to a chip company and ask, "what can you build for us?" This is because there is a lot of work involved in doing a new architecture. So why would other companies want to make such a large investment in a custom CPU design even if they don't have to build an ASIC to get it built? I don't want to doubt that you are seeing the results you claim. But Xilinx has publicly said that the Sp3 is not as fast as the VII/VIIpro. So unless you can get the same or better results targeting the Virtex family, I suspect your results are anomolous. But keep us informed, this is very interesting. -- Rick "rickman" Collins rick.collins@XYarius.com Ignore the reply address. To email me use the above address with the XY removed. Arius - A Signal Processing Solutions Company Specializing in DSP and FPGA design URL http://www.arius.com 4 King Ave 301-682-7772 Voice Frederick, MD 21701-3110 301-682-7666 FAX
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