Re: [casper] large green FFTs
Last time I tried this on 10.1 with a 2^20 FFT design, the Matlab froze for a long time and I ended up closing the program. On Mon, Jul 27, 2009 at 8:31 AM, Laura Spitler laura.spit...@gmail.comwrote: Hello everyone, What's the status of large green FFTs, where large is greater than 2^11? Can they now be reliably synthesized in 10.1? Thanks, Laura
Re: [casper] large green FFTs
I compiled an 8192 channel FFT the other day no problems. But I have not yet tried it on actual hardware. Jason On 27 Jul 2009, at 08:41, Jason Zheng wrote: Last time I tried this on 10.1 with a 2^20 FFT design, the Matlab froze for a long time and I ended up closing the program. On Mon, Jul 27, 2009 at 8:31 AM, Laura Spitler laura.spit...@gmail.com wrote: Hello everyone, What's the status of large green FFTs, where large is greater than 2^11? Can they now be reliably synthesized in 10.1? Thanks, Laura
Re: [casper] large green FFTs
hi jason zheng, as you probably know, a 2^20 point FFT won't fit on an fpga, as it requires 2^19 times 36 bits of memory. the largest single fft that we've ever used on an FPGA is 2^15 points. to implement higher resolution spectrum analyzers one needs to implement analysis in two stages (course PFB, followed by fine FFT). for example see the 128 million point spectrometer block diagram at http://seti.berkeley.edu/jplsetispec/block_diagram.pdf best wishes, dan Jason Zheng wrote: Last time I tried this on 10.1 with a 2^20 FFT design, the Matlab froze for a long time and I ended up closing the program. On Mon, Jul 27, 2009 at 8:31 AM, Laura Spitler laura.spit...@gmail.com mailto:laura.spit...@gmail.com wrote: Hello everyone, What's the status of large green FFTs, where large is greater than 2^11? Can they now be reliably synthesized in 10.1? Thanks, Laura
Re: [casper] large green FFTs
Dan, Btw, thanks for the pointer, was the 128M point spectrometer realized using Casper library? ~Jason On Mon, Jul 27, 2009 at 9:13 AM, Jason Zheng jxzh...@gmail.com wrote: Dan, I fully understand the memory limitations. Just to clarify, it was just a experiment on the capabilities of the Casper FFT library, not intended to actually implement on FPGA. What prompted me to do this experiment was a rumor that the Casper library could support multi-million point FFTs. ~Jason Zheng On Mon, Jul 27, 2009 at 9:09 AM, Dan Werthimer d...@ssl.berkeley.eduwrote: hi jason zheng, as you probably know, a 2^20 point FFT won't fit on an fpga, as it requires 2^19 times 36 bits of memory. the largest single fft that we've ever used on an FPGA is 2^15 points. to implement higher resolution spectrum analyzers one needs to implement analysis in two stages (course PFB, followed by fine FFT). for example see the 128 million point spectrometer block diagram at http://seti.berkeley.edu/jplsetispec/block_diagram.pdf best wishes, dan Jason Zheng wrote: Last time I tried this on 10.1 with a 2^20 FFT design, the Matlab froze for a long time and I ended up closing the program. On Mon, Jul 27, 2009 at 8:31 AM, Laura Spitler laura.spit...@gmail.commailto: laura.spit...@gmail.com wrote: Hello everyone, What's the status of large green FFTs, where large is greater than 2^11? Can they now be reliably synthesized in 10.1? Thanks, Laura
Re: [casper] large green FFTs
Btw, thanks for the pointer, was the 128M point spectrometer realized using Casper library? yes, details of this spectromter are at: http://casper.berkeley.edu/wiki/SETI_Spectrometer best, dan ~Jason On Mon, Jul 27, 2009 at 9:13 AM, Jason Zheng jxzh...@gmail.com mailto:jxzh...@gmail.com wrote: Dan, I fully understand the memory limitations. Just to clarify, it was just a experiment on the capabilities of the Casper FFT library, not intended to actually implement on FPGA. What prompted me to do this experiment was a rumor that the Casper library could support multi-million point FFTs. ~Jason Zheng On Mon, Jul 27, 2009 at 9:09 AM, Dan Werthimer d...@ssl.berkeley.edu mailto:d...@ssl.berkeley.edu wrote: hi jason zheng, as you probably know, a 2^20 point FFT won't fit on an fpga, as it requires 2^19 times 36 bits of memory. the largest single fft that we've ever used on an FPGA is 2^15 points. to implement higher resolution spectrum analyzers one needs to implement analysis in two stages (course PFB, followed by fine FFT). for example see the 128 million point spectrometer block diagram at http://seti.berkeley.edu/jplsetispec/block_diagram.pdf best wishes, dan Jason Zheng wrote: Last time I tried this on 10.1 with a 2^20 FFT design, the Matlab froze for a long time and I ended up closing the program. On Mon, Jul 27, 2009 at 8:31 AM, Laura Spitler laura.spit...@gmail.com mailto:laura.spit...@gmail.com mailto:laura.spit...@gmail.com mailto:laura.spit...@gmail.com wrote: Hello everyone, What's the status of large green FFTs, where large is greater than 2^11? Can they now be reliably synthesized in 10.1? Thanks, Laura
Re: [casper] large green FFTs
I suggest you try the 10.1. I no longer use the 7.1 libraries and they're not under active development like the 10.1 libraries. There have been numerous fixes to the 10.1 FFT to fix these and other problems. There are still timing problems with 10.1, but the FFT is working AFAIK. My compile was 8192 channels (16384 channel) at 200MHz. Jason On 27 Jul 2009, at 11:02, John Ford wrote: I compiled an 8192 channel FFT the other day no problems. But I have not yet tried it on actual hardware. We had trouble building large ones that meet timing. Randy modified the library to fix the problem. I'm thinking these were 4096 channel (8192 point) real 200 MHz IP rate builds. It works for us... This is using Xilinx version 7.X stuff. I don't know if the same fix would help the 10.1 version. John Jason On 27 Jul 2009, at 08:41, Jason Zheng wrote: Last time I tried this on 10.1 with a 2^20 FFT design, the Matlab froze for a long time and I ended up closing the program. On Mon, Jul 27, 2009 at 8:31 AM, Laura Spitler laura.spit...@gmail.com wrote: Hello everyone, What's the status of large green FFTs, where large is greater than 2^11? Can they now be reliably synthesized in 10.1? Thanks, Laura
Re: [casper] large green FFTs
I suggest you try the 10.1. I no longer use the 7.1 libraries and they're not under active development like the 10.1 libraries. There have been numerous fixes to the 10.1 FFT to fix these and other problems. There are still timing problems with 10.1, but the FFT is working AFAIK. Yeah, well, we will release our machine to the public on Thursday, and the team would kill me if I suggested we moved to a different tool set right now. It takes a few weeks to run all the tests we have to run to verify that things work right for production use... We do have a 10.1 toolset installed, but it doesn't quite work yet. We have a couple of ROACH boards in the mail to us, so I guess we'll be migrating to 10.1 soon enough! John My compile was 8192 channels (16384 channel) at 200MHz. Jason On 27 Jul 2009, at 11:02, John Ford wrote: I compiled an 8192 channel FFT the other day no problems. But I have not yet tried it on actual hardware. We had trouble building large ones that meet timing. Randy modified the library to fix the problem. I'm thinking these were 4096 channel (8192 point) real 200 MHz IP rate builds. It works for us... This is using Xilinx version 7.X stuff. I don't know if the same fix would help the 10.1 version. John Jason On 27 Jul 2009, at 08:41, Jason Zheng wrote: Last time I tried this on 10.1 with a 2^20 FFT design, the Matlab froze for a long time and I ended up closing the program. On Mon, Jul 27, 2009 at 8:31 AM, Laura Spitler laura.spit...@gmail.com wrote: Hello everyone, What's the status of large green FFTs, where large is greater than 2^11? Can they now be reliably synthesized in 10.1? Thanks, Laura
Re: [casper] large green FFTs
I should have provided some background for my question. I have a basic spectrometer model for an iBOB that uses 10.1. It works perfectly for FFT sizes up to 2k. Over the weekend I tried 4k and 8k, and both resulted in incorrect spectra. Timing was fine. So has anyone used 10.1 to compile a model with a FFT larger than 2k and successfully tested it in hardware? Laura On Mon, Jul 27, 2009 at 2:06 PM, Jason Manleyjasonman...@gmail.com wrote: I suggest you try the 10.1. I no longer use the 7.1 libraries and they're not under active development like the 10.1 libraries. There have been numerous fixes to the 10.1 FFT to fix these and other problems. There are still timing problems with 10.1, but the FFT is working AFAIK. My compile was 8192 channels (16384 channel) at 200MHz. Jason On 27 Jul 2009, at 11:02, John Ford wrote: I compiled an 8192 channel FFT the other day no problems. But I have not yet tried it on actual hardware. We had trouble building large ones that meet timing. Randy modified the library to fix the problem. I'm thinking these were 4096 channel (8192 point) real 200 MHz IP rate builds. It works for us... This is using Xilinx version 7.X stuff. I don't know if the same fix would help the 10.1 version. John Jason On 27 Jul 2009, at 08:41, Jason Zheng wrote: Last time I tried this on 10.1 with a 2^20 FFT design, the Matlab froze for a long time and I ended up closing the program. On Mon, Jul 27, 2009 at 8:31 AM, Laura Spitler laura.spit...@gmail.com wrote: Hello everyone, What's the status of large green FFTs, where large is greater than 2^11? Can they now be reliably synthesized in 10.1? Thanks, Laura