Kristian Zarb Adami wrote: > We have terminated two inputs on the 1GSA/s ADC board and run them > through an FX-correlator. The left hand panel shows the cross-power > spectrum after an integration of 20s. > > I would expect the noise floor to keep decreasing, but for some > reason it doesn't and remains at ~ -110dBm with this structure. > > When we capture raw-ADC data, perform the same thing floating point > we don't see such structures, which indicates that this might be a > fixed-point issue.
Hi Kris, It sounds like your noise floor is bottoming out at the quantization noise floor that you'd expect with 18 bit fixed point data. You won't get better than ~110dB by integrating these further. Not sure what kind of floating point you're using, but single precision floats use a 24 bit mantissa. If you integrate these for long enough, you'd probably bottom out at ~145dB. If 110dB isn't good enough, you need to use more bits. Each single-bit increase should lower your quantization noise floor by 6dB. (Or so i'd guess. Someone who's actually studied DSP should correct me if i'm wrong.) Billy
