I've done a 256 kpoint biplex fft in a virtex 7. You could probably do 128kpoint in a v6. By combining biplex channels this probably means a 512 kpoint fully on-chip fft, with no pfb and a little ip development for the direct form part to work.
You can get an idea of how much a pfb will cost by taking (pfb_len*[ntaps-1])~=effective_fft_len which applies for memory utilization. (Which should be your limiting factor) This estimate breaks down for high bandwidth applications, and you'll need to do a few things on your fft to be that efficient. I can talk more if you're interested. On Jan 20, 2014 9:43 PM, "Jason Manley" <[email protected]> wrote: > There was a bug preventing FFTs over 2^16 being compiled. I haven't > retried this after Andrew's mods but hopefully this is fixed. You will run > out of BRAM trying to compile very large PFBs. It's easier to use the 2-D > approach that Dan describes, if you can accommodate the weird > spectral/channel artefacts that this introduces... > > I know of at least one spectrometer built as Dan describes, using a > 256x4096 channel PFB (i.e. 1M channels) over ~1GHz BW and there were plenty > of FPGA resources left for building larger. This design was limited by > external memory, which was being used for other things too (you'll need a > big VACC!). Also to note, the readout speeds become rather fast at these > high resolutions. > > Jason Manley > CBF Manager > SKA-SA > > Cell: +27 82 662 7726 > Work: +27 21 506 7300 > > On 21 Jan 2014, at 7:13, Dan Werthimer <[email protected]> wrote: > > > > > > > hi gerry, > > > > > > we haven't tried this, but i think the largest spectrometer you > > could fit on a roach2 is 256M points, implemented by a 16K point FFT, > > followed by DRAM based corner turn and twiddle factors, > > followed by another 16K point FFT. > > > > if you have this many channels in your correlator, > > you also be running up near the correlator X engine memory limits: > > > > for instance, if you cross correlate in a Titan GPU, then you only have > > 5 or 6 GB of memory on each GPU card. > > > > let's assume you have a max of 32 GPU's for your X engine. > > > > then max frequency channels = > > > > 32 GPU's x 6GB/GPU x 42^2 baselinepols x 4B/baseline > > > > = 435M channels max for 32 GPU's (round down to 256M max channels) > > > > > > if you cross correlate in a CPU (eg: DiFX) then you can have more memory, > > but you'll need a lot more CPU's to keep up with the data rate, so CPU's > > won't help. > > > > be wary of readout rate too - that's a lot of data to read out : > > > > 256M channels x 42^2 baselinepols x 4B = 1 TB every integration > time > > > > > > > > best wishes, > > > > dan > > > > > > > > > > On Mon, Jan 20, 2014 at 7:53 PM, Gerry Harp <[email protected]> wrote: > > Hi > > > > Just for fun, how large of an FFT (filter bank) can fit into one of the > Roach# boards? Has anyone ever successfully compiled a filter bank with > length 2^17? We're interested in building a relatively narrow-band > correlator so we need lots of channels. Any experience at large lengths or > educated guesses are welcome. Also, how fast did it go? Possible to keep up > with 100 MSPS? > > > > It is proposal time, once more... > > > > Thanks > > > > Gerry Harp > > > > > > On 1/17/2014 11:56 AM, Dan Werthimer wrote: > > > > > > > > > > Dear Casper Collaborators, > > > > > > We hope you can attend this year's Casper Worshop > > > > in Berkeley, California > > > > June 9 throuh June 13, 2014 > > > > > > > > > > We'll have more information later about registration, > > travel, abstracts, etc, but for now, please reserve these dates. > > > > > > Hoping you can participate, > > > > > > Dan and the Scientific and Local Organizing Committees > > > > > > > > > > > > > > > > -- > > ---------------------- > > Gerald R. Harp, Ph.D. > > Director, Center for SETI Research > > SETI Institute > > > > > > > > >
