Why does noise floor disappear depending on n bits?

[macy]

In analyzing a system with pure tone signal that is 2.36 Vpk at 31.250 Hz,
synchronous sampling at 1000 S/s data acquisition, digitized to 24 bits
within +-2.5V.  Also, there is approx 5 nV/rtHz added to the signal.

When a 8192-point FFT is done, some strange things happen when plotting
the resulting FFT in dB scale.

The result looks about right, but digitize using 23 bits, again about
right notice the noise floor is slightly increasing. 22 bits, still ok, 21
ok, and 20 ok.

BUT! when I digitize with 19 bits, the noise floor of the FFT drops to
zilch, like -300 dB?!  Then 18 bits ok again, but then at 17 bits, again
with the no noise floor.

What am I doing wrong here?  Anyone have an explanation?

section of code looks like

f=31.25;A=2.36;t=[1:8192]/1000;
sig=A*cos(2*pi()*t);
vnoise=5e-9;
vn=sqrt(500)*vnoise*randn(1,8192);
sn=sig+nv;
n=24;Vmax=2.5;Vmin=-2.5;
dV=(Vmax-Vmin)/2^n;
s=dV*round((sig1n-Vmin)/dV)+Vmin;
b=sqrt(2)*fft(s)/8192;
blog=20*log10(abs(b(1:4096)));
blog=blog-max(blog);

Any idea why certain n-bits causes the expected noise floor to simply
disappear?


Robert