Check the memory chip sockets for broken solder joints.

Maybe a video mode change? from 3 to 10h? 720×400 -> 640×350 or other way around. That would require changing video clock.

The fact that the screen corruption is randomly different between resets, but consistent within a session hints at some logic issue, but not physical problem.
I touched, pressed, and slightly bent things around - if this affects appearance on screen that is a good indication of loose connections. Not the case with this card.
The memory is soldered through through-holes, no sockets. I reduced it to 512Kb only atm.

mov ah, 10h
results in garbled mode 3 screen.
Not sure what this means.

pshipkov wrote on 2024-09-12, 17:09:

Everything looks good until the memory count completes and the table-screen is about to appear.

That's the moment when some chipset settings from the BIOS get applied. Possibly that's also the moment in which the system switches from using the actual video BIOS chip to using a shadowed BIOS. For further troubleshooting, try disabling shadowing of the video BIOS, try disabling L2 cache, and try more conservative memory timings.

Was about to reply "already tried all that" but went one more time over the BIOS settings and it does look like i didn't check video shadowing before. Or i probably did, but in combination with other stuff which masked the solution.
So, disabling VIDEO SHADOWING only fixes the problem. Glad i double checked before responding.

Now, disabled VIDEO SHADOW is impractical. 90ns latency for EPROM bound calls is just not cutting it.

Color me surprised. How would vbios shadowing screw with video output while not crashing computer?

First time seeing this kind of stuff myself.