This is not a bug in Safbench, but: if you're overclocking your video card,
the results might be unstable (+-0.1-1.0 MB/s) if using option 'V' with write
gathering/combining enabled or just option 'F' (at least on my 6x86).

'B' and 'P' in Safbench doesn't give the REAL MHz-rate at which the bus
is running; it just displays the number of clock cycles during which the
processor's external memory bus was in use while reading data from main
memory. 101.000 MHz means that the bus was busy about 101.000.000 cycles
in one second. This value can't be larger than CPU's MHz-rate.
Performance monitoring event 18h is used on Pentiums and Pentium MMX's and
62h on PPros and compatibles to calculate the BUS MHz-rate
(if ALSO 'C' is used, 22h is used for Pentium Pros (option 'P')).

Win-bugs I've seen: once in a DOS-session under Win95 time slowed down;
timer ticks got generated three times less than normally (reported times
were 3x worse). Once Safbench seemed to hang in "Delaying..." --- but
actually timer ticks didn't get generated at all. I verified it with SoftICE.
After exiting and restarting 4DOS-shell it worked OK.
And these were only Safbench- or Qsort32-related Win95-bugs... ;)