Next-Generation "Dream" Pulsar Machine Notes
            Scott Ransom
             4 Sept 2006

General requirements:
  --  Publicly available from outset
  --  Architecture should be based as much as possible on
      commodity off-the-shalf components (Note that I consider
      FPGAs off-the-shelf.  Also, custom boards based on FPGAs
      and "standard" DSP chips is also OK.)
  --  Independent machine (except for back-of-backend, see below)
      i.e. does not use GBT spectrometer for ACFs
  --  Total system is not a substantial power or AC hog
      (this precludes us from doing everything in software
      with ~100 fast servers)
  --  High dynamic range sampling of baseband signal
  --  Possibility for active RFI mitigation (at least for narrow
      band signals) on board would be great...
  --  Dual polarizations in voltage; Full Stokes in power
  --  High linearity (in general, it would be great if the ADCs
      could be pushed as far upstream as possible -- maybe
      even into the frontend(s)

Input:  
  --  "real" sampling baseband signals of BW from ~5 MHz to ~1000+MHz
  --  Digital samplers are part of the system (as above, pushed
      as far upstream as possible)
  --  Sampling should be at least 8-bit and preferably 12-bit
      (maybe 12-bit for BW<100MHz and 8-bit for BW>100MHz?)

Monitor and Control:
  --  Ability to tap in to each section of system to have "scope"-
      like access to ADC, FFT, PFB performance
  --  Realtime passband and calibration monitoring
  --  Viewing of firmware/hardware settings etc

Output:
  --  Output is sent to one or more data-handling machines
      --  Optimally, the data would be packetized over 1- or 10gigE
      --  The number of backend servers should be small (<~ 8)
          in order to minimize power and AC requirements
      --  Data handling machines handle formatting (i.e. FITs), 
          folding, downsampling, and writing of data to disk
  --  Two types of basic output
      --  Incoherent filterbank (Polyphase or similar)
          --  1 IF, 2 IF (unsummed or summed), or Full Stokes
          --  Should allow between 256-4096 channels
          --  Several BWs should be possible
              (~5MHz, ~20MHz, ~64MHz, ~200Mhz, 600MHz, 1GHz+)
          --  Hardware output will optimally be "simple" (i.e. in order)
          --  Several quantization levels should be allowed
              4-, 8-, 16-bit ints?  (this could happen in hardware 
              or by processing in the backend machines)
          --  Backend machines will write data in FITs format
      --  Coherently dedispersed subbands (processed in parallel)
          --  Total coherent dedispersion capability (to be handled
              in hardware) should cover at least 600 MHz (i.e.
              matched to clean chunk of S-band) for "high" DM MSPs 
              (i.e. DMs <~ 300 pc cm-3).  Optimally, BWs of up to 
              ~1GHz would be great...
          --  Number of subbands should be tunable (4-64 ?)
          --  Subbands are folded in realtime or combined,
              downsampled and written to disk by backend machines
          --  Folding of multiple pulsars in hardware would be
              optimal, however, for many pulsars, we may need to 
              use computer backends (Note:  folding should likely
              use the new TEMPO2-style polycos)