As I have been experimenting with "newer" hardware to run old OS, I realized that in most cases, the issue was not much to get the OS to run on the hardware but to have the OS correctly address the hardware through the drivers.

It may sound something pretty obvious, but it is not necessarily like that. For example; for a CPU-based/bound game, that leverage on how fast a CPU is; you can brute force things around a newer CPU even if there are no optimizations. This of course is not entirely applicable, as we all know that modern CPU still use 3DNOW, SSE, MMX and all the other CPU extensions from the past for most part. So in the end you can still emulate a CPU, and brute force it on a modern CPU using the raw power of a modern CPU to simulate/emulate everything on it.

But for GPUs the issue is quite different.

For GPU, even going from DX3 to DX7 you lose optimizations and a game may lose a lot of performances or not run at all; and assuming you install DX runtimes, if the video card drivers cannot leverage on those instructions because the card is too new, you end up getting worst performances compared to older hardware that is less powerful for example. I incorrectly assumed that I could bruteforce everything with a newer machine, but data show me something quite different, so I want to bring up what I found, and hear from you if you can work around this... Or if in the end is not worth waste time as it is a non-problem.

My scenario: I collected different hardware to install W98 and XP; then I installed 3DMark to get some numbers and then I ran some games to see how realistic the 3Dmark results were.

The machines:

1) Dell Wyse 7020 thin client, with a 32 GB SSD and 4 GB of RAM. CPU is AMD APU GX-420CA
2) HP Elitedesk 705 G1 , with a 32 GB SSD and 8 GB of Ram. CPU is AMD A8 Pro 7600B
3) Compaq Pro 6305, with a 32 GB SSD and 8 GB of Ram. CPU is AMD A8-5500
4) Chuwii micro computer, with 128 GB SSD and 16 GB of RAM. CPU is Intel Celeron J4125

First round: W98. They were able to boot in W98 after installation except the Chuwii. Although I could not find any driver for most of the hardware, so I was running at 640x480 with no acceleration nor audio; which made me think that was useless to even spend time at that point doing anything further.

Second round: XP
As before , they all installed and booted in XP except the Chuwii. Clearly that computer is not able to boot unless it is UEFI, so not useful to run anything old as far as OS.

After installing drivers. this is what I got from 3Dmark

3dmark 2000

hp 13430
wyse 12926
Compaq 35269

3dmark 2001

hp 12638
wyse 16050
compaq 30951

3dmark 03

hp 10405
wyse 10971
compaq 23787

3dmark 05

hp 5653
wyse 6869
compaq 13382

These results are with the onboard GPU for these CPU; they are all AMD APU, so they use the Radeon drivers/Catalyst software with default settings. I was able to find a video card to use on the Compaq; as that is the only computer that has a slot for a discrete GPU; and to my surprise, the performances with it were worst actually

3dmark 2000

compaq with GPU 11956
Compaq 35269

3dmark 2001

compaq with GPU 15221
compaq 30951

3dmark 03

compaq with GPU 10631
compaq 23787

3dmark 05

compaq with GPU 6798
compaq 13382

While looking at this data there was something it was not making sense, so I looked at the fabrication time for each of these APU and the discrete GPU and it still doesn't make sense.

The Compaq is the best of the bunch; it is running from a GPU that is a Radeon 7560D IGP on the AMD A8 APU; so it is integrated graphics; and that is twice as fast as both the HP and the Wyze. And twice as fast as a discrete Radeon HD7570 with 1 GB of dedicated VRAM... The delta between 3dMark tests is the same; where the HP is the weakest of all, followed by the Wize; then the Compaq with the discrete Radeon HD7570 and the best one is the onboard 7560 from the A8 APU.

Is this something that is actually realistic? I am using the same exact OS version for XP; the drivers are the last used by Windows XP from the AMD website (so latest Adrenaline official for Xp 32 bit); and even changing driver version I noticed no relevant changes; as the hierarchy was basically the same.
I would assume that the drivers are basically doing the best work on the A8 APU at this point? Because that APU is not so much more powerful than the others in the other machines to justify twice as much in the final score results to be honest. And I found nothing that would make the other machines (or the Compaq using the discrete GPU) to go twice as fast either; so this is really baffling me.

The A8-5500 on the Compaq was released in 2012; has 4 cores and 65W total; the Radeon 7560D is close in performances to a GT430 (which is 20% faster)
The A8 Pro 7600B in the HP was released in 2014; same 4 cores as the A8-5500 but it should be faster. The Radeon R7 on it should be 20% faster than the GT430, making it 40% faster than the GPU on the A8-5500... Instead the latter is twice as fast as the former.
And the Wyze is running on a Radeon HD8400E, released in 2013; which is 300% slower than a GT430; so the difference between the 2 previous devices and the Wyze should be like an abyss... And instead the HP and the Wyze score almost the same. As if either the Wyze with its low tier 25W CPU is basically doing the best it can, versus the other 2 APUs which should score way more, being faster than the Wyze.

To me all of this seems to be related to the fact that most of these issues are all trelated to the drivers and their versions; I suspect that in the end some hardware may be supported better than others because they are maximizing to the fullest the hardware, while in the case of the newer hardware it is just barely used, so the performances are just barebone.

Does anyone have a technical explanation for this illogical number outputs? I ran the test multiple times so I know they are correct, and I want to learn more about what is happening here because it is fascinating as hell.

And if you think that it is a fluke of 3dMark; the same results are show when running games... The Compaq goes twice as fast indeed.