interesting, I thought the AMD 386 where 100% compatible.... maybe the fpu is causing the probs somehow?

Since the AMD 386 processors were licensed from Intel, they should be 100% compatible. AMD 486 CPUs are another story, because they have a different CPUID, which causes problems with some programs, but on 386 the CPUID feature didn't exist yet.

In this case it's rather difficult to find out what causes the problem, as you can't change the CPU without changing the board.

About Optic Nerve: It's not really worth bothering, this one is almost impossible to get running on any machine. Many years ago I watched it halfway through on my 386SX, then it crashed and never worked again.
Someone on USENET suggested to make this into a competition at scene parties: Get Optic Nerve running on real hardware 😁
Btw, it works without problems in DOSBox 😉

I could probably help with the other three demos, if you post more info about your machine (video card, sound card, mainboard chipset, DOS config, etc...)

I've completed the test:

Dragnet:
intel: Works with EMM386
amd: EMM386 = Error 12 after intro (Stack-Segment Fault), works correctly with only himem.sys

Optic Nerve:
intel: works with 386 patch (slowly), doesn't load with memory manager
amd: reboot while loading

Cronologia:
seems to be the same problem with Dragnet after the 1st part
crash on amd while loading the 4th part

It seems to be a difference between virtual mode on intel and amd

Here is my configs:
chipset: intel = PC CHIPS TCL14L100 / AMD = ALI M1217-40
ram: 16MB
gfx: Trident TVGA9000C 512kB vesa 1.2 (sometimes crash during its initialisation on intel mobo)
sound: SB16 ASP CT2290
scsi: AHA-1540CP
nic: 3C509B (sometimes crash while loading its drivers on intel mobo)
i/o: on intel mobo, I was using DTK PTI-217 but I had to replace it with M5105A4 because it prevents amd mb to post
OS: dos 6.22
Turbo: intel = 25MHz->8MHz / AMD = memory WS

amd mobo has jumpers for 20/25/33/40MHz and socket for cristal to make fpu running asynchronously (IIRC only intel fpu can do this)

HIGHSCAN/I=F000-F7FF only work on amd mobo

if the amd cpu is just a licensed clone why cant you use the same board ?

ps: could the 16mb of ram be a problem no-one had 16mb back in the 386 days my first pentium pc only had 8mb

386sx are soldered, there weren't any socketed version.

I found from this page

The only noticeable difference I found was that they didn't prefetch instructions the same way, and when you disabled the external cache, you could notice a different pipeline stall depending on instruction alignment.

I've got an Am386DX-40 mobo
chipset UM82C482AF, empty fpu socket, 256kB cache, MR BIOS
CTRL ALT - 8MHz
CTRL ALT + 40MHz

DTK PTI-217 works with it.

scsi controlers give me some troubles with this mobo:

AHA1542:
crash during its initialisation when its bios is shadowed
hang on disk access when hw slowdown enable
I can run Optic Nerve

AHA1520:
Shadow and slowdown work
Optic Nerve make reboot while loading

I'll try to test the other demos with and without cache but I dont have any i386DX to compare against.

Sometimes, Optic Nerve crash after Gouraud shading on this mobo
I think it when it is run after cold reboot with cache disable