A problematic rev3 A500

[jsan]

Hi!


I've spent the last few months trying to diagnose a rev3 A500 to no avail. I am sending this message as a last resort to see if we could crowdsource the solution to this mistery.


Before describing the issue, a quick note to let you know I have:
- Replaced all IC sockets in the board including ram, logic, gary, agnus and CIAs. Pretty much everything except paula and denis.
- Tested with different ICs: all logic ICs are brand new, and I've tested a second set of all other ICs (68k, agnus, paula, denis, gary, CIAs)
- All lines have been tested for continuity, data and address bus, as well as any other signals between ICs in the board. I can't be 100% sure that I tested all of them, but I was pretty methodical in the process so the chances I miss one are low.
- The results shown are 100% repeatible, not a single bit changes between different runs.
- Only 512Kb of chip memory is present in the computer.
- Kickstart 1.2 and 1.3 won't ever get to the insert disk screen, a guru error is returned in most occasions.



So, there seems to be a problem in the access to memory. However, it's not a straighforward issue. Diagrom 1.2.1 completes the basic chipmem check without any issues. And I can pass pretty much all the other tests, except a couple of the graphical tests (scrolling being one of them). Diagrom 1.3 beta introduces a better chip memory testing methodology, where each address of memory stores it's own address value. This test fails miserably.


In my testing, I have found the Manual Memory Testing to be of great help. I can explore the chip memory freely, and I've found the strangest issues. Look at $50000:


All positions in this area should contain their address as the value, yet instead of 50000, 50004, 50008 what I'm getting is 40000, 40004, 40008. I've found this is repatable for addresses starting with odd numbers (10000, 30000, 50000). For addresses starting with even number (20000, 40000) it looks fine, at least the very first block.


Even more interesting, if I manually edit some positions in this memory area, when I refresh the memory values (pressing R) things seem to not be stored properly. For example, I filled with 11s a bunch of positions in the block:


Just after refreshing, the result is:

A bunch of 1s have turned into 0s. Always on the LSB, but it's not 100% of them. So it's not a sticky bit. And note that sticky bits are perfectly captured by Diagrom tests.


To make things even better, if I refresh a second time, things seem to change again:


Even more 1s turned into 0s.


Now, it gets even better. Just changing one position of memory, setting $4ffe0 to $07 produces changes in another 3 different memory positions: $50044 (changes the value from $44 to $45), $50094 ($94->$95), $500E4 ($E4->$E5). It's likely that there are additional memory positions that get changed, but these are the ones visible in the screen.


It doesn't seem plausible that editing a single position of memory alters the content of multiple memory bytes. So everything seems to point to an issue in the address bus: retrieving the information for $50094 does not return such address, but another. However, the fault should be in the board given the tests are identical with a complete different set of chips.

I'm reaching a dead end, and seriously considering just keeping the chips as spares for my other machines. But it makes me sad to accept this unfortunate fate for this machine. Any help will be greatly appreciated!

Thanks for reading until the end, I hope we can all solve this mystery together! Don't hesitate to ask me for any additional information.


Cheers!


Jose