Hi John,
regarding your request for a dual 1M point spectrometer:
As you know Terry ported PASP to Roach (Packetized Astronomy Signal
Processor).
PASP digitizes a pair of 400 MHz bandwidth signals,
breaks these signals into 2^k sub-bands using a PFB,
and then packetizes and transmits the data via a switch to bank of
CPU/GPU's.
The switch isn't required if you have four or less CPU/GPU's.
Terry and Mark have written code which receives these packets and
forwards them to a GPU which does FFT's, compute's power spectra
etc. (very similar to Paul Demorest's code). Our Japanese colleagues
have developed a fast CUDA spectrometer code that can be spread
over multiple GPU's. This code is programmable in the FFT size,
so I think you could configure it for 1M channels.
Instead of building your spectrometer from two stages - course and fine
(PFB in FPGA, FFT in GPU), it might be better
to send raw 400 MHz time domain data into the GPU's (set k=0),
so that you don't introduce scalloping from the PFB.
Terry has a design that does this, and Matthew Bailes and his colleagues
have developed back end CPU/GPU code and have been using it at Parkes.
Best Wishes,
Dan
On 5/26/2010 5:41 AM, John Ford wrote:
Hi all.
I know folks are working on a hybrid spectrometer using roach/ibob and
GPU's. I am in need of a 1M channel, 2 pol, 400 MHz spectrometer. Anyone
offer up their design as a possibility? I think GUPPI maybe could do it,
but I wonder if anyone else has gotten anything going. Or could you build
such beast with just a roach or 2?
John
