Red wires are straight traces and those are correct.
White follows a trace that was like that before removing fake chips.

-

After soldering a defective joint, cache is now detected.
I have to put TAG ram to 8+0 bits, if 7+1 it won't work with stock bios.
Write back or through works.

Using Phil's boot menu cache check:
extended memory: cacheck loads fast and detects L1 + L2
expanded: uber slow loading + cpu is in V86 mode warning and detects L1 + L2
conventional: says mb #x or y or z looks not cached error

I get mb no. not cached, don't get why tho.

The performance bump is noticeable.
Quake benches almost doubled at any resolution, and others too.

Tried with 1995X bios chip. Tomorrow I'll try with the one it came from.

BIOS 1995X is so far the best to provide performance.

I suspect 1995D to have some hardcoded settings that prime some aspects in ram performance rather than the whole system of L1+L2+Ram.

In WT mode for cache they are on par as long as you don't fiddle with other parameters, as 1995D won't allow those and you'll end up with a slower system.

Cpus I tried can't work in WB, no idea why but L1 WB or WT give the same results.

Went so far and that's it. I'll post a few speedsys and cachechk pics.

Bottom line: if you have this model with fake chip, if pins go to ram and chipset most probably it'll work with real L2.
Update to 1995X BIOS version.
---

IBM 5x86 100 mhz is a nice upgrade.