hi david,
there are a few casper instruments that are similar to what you are
suggesting.
1) some of the VEGAS spectrometer modes use a roach to digitize the data and
send almost raw adc samples to a GPU where the spectroscopy is done.
2) some of the GUPPI pulsar machines use casper hardware to only digitize
the data - the GPU's do the DSP.
3) the astronomy signal processor sends raw adc data to a bank of cpu/gpus
or breaks it into sub-bands and sends each frequency band to a
different cpu/gpu.
4) the GMRT correlator digitizes data on PCIe boards (some versions use
roach boards)
and then they have options to do the processing either on CPU's, GPU's
or FPGA's.
best wishes,
dan
On Mon, Dec 5, 2011 at 3:09 PM, David Forbes <dfor...@email.arizona.edu>wrote:
> CASPERites,
>
> I had a crazy idea the other day, and wanted to see if it's crazy enough.
> Hence, I post it here.
>
> The idea is to build a spectrometer using off-the-shelf PC hardware, with
> only a custom data acquisition board. All the DSP would be done by the PC.
> The NVidia Tesla GPU board is no slouch at DSP, having 512 CUDA processors
> clocked at ~1GHz. It also plugs into a regular PC motherboard.
>
> A person could make a relatively low-cost PCI Express card with a fast ADC
> and a smallish FPGA with just a data path from the ADC to the PCI bus. The
> data would blast into the GPU and be processed by those 512 CUDAs,
> resulting in output spectra calculated in the usual way.
>
> One would have to select a motherboard with a proper PCI Express switch
> that would actually allow the data to flow at full 16x rate from the ADC
> board to the GPU.
>
> I wonder just how this approach would compare to the ROACH II, for example.
> It seems that the cost would be ~4x lower and the performance within a
> factor of two. It also reduces the complexity of the non-mainstream
> hardware, allowing most of the budget to be spent on products with
> competitive sales channels.
>
> Thoughts?
>
> --David Forbes
> Arizona Radio Observatory
> Tucson, AZ
>
>
>
