Here's what i get on a 5000/150 with 1.4 (distribution):

Cuisinart> xntpdc -c sysinfo
system peer:          Gatekeeper.DSG.Stanford.EDU
system peer mode:     client
leap indicator:       00
stratum:              3
precision:            -17
root distance:        0.00955 s
root dispersion:      0.02899 s
reference ID:         [171.64.79.3]
reference time:       baf4e0cf.54e52000  Tue, May 25 1999  1:41:51.331
system flags:         bclient auth monitor pll stats kernel_sync 
frequency:            0.000 ppm
stability:            423.399 ppm
broadcastdelay:       0.003906 s
authdelay:            0.000030 s

(I used to be able to get the precision down to -18; should get there
again if we can fix thecache conflicts and/or I build a non-GENERIC kernel).

5000/240s are also very robust, they use the TC cycle-counter to
interpolate ticks.  I'd guess either you're losing interrupts, or the
calculation of tick length is busted.

Do you see the same thing with a 1.4 kernel?