Sysinfo 3.24 shows different speed results if you start it once, then move a few windows or do "avail flush" and restart it. I guess that the speed test code is depending very much on cache line alignment: If the core routine can take advantage of a full burst (4 longwords), it's faster and shows a higher value, although the machine is identical.
Is it possible to re-align the routine in fastmem to an address that has the lower 4 bits =0? Or even "configurable", so the routine is shifted in memory by 1-3 long words and the four values of four speed runs are shown in some debug window? Would be great to have the speed rating reproducable.
Another tool I use in accelerator development is "Bustest". Sometimes I think that an additional "speed" button on a memory list entry would be nice. Well, if you want additional items on your todo list ;-)