A recent post by @mpe here reminded me of something i started before, but never finished properly - basically, try to find the best possible use of a slightly modified P24T (POD83) running at 100MHz.
The reason i left the job for later was the overall instability of the CPU with most Socket 3 motherboards.
Only few handled it properly. When it worked, performance significantly increased in several areas.
So, i replaced the 160MHz AMD processor with POD100 in one of the 486 PCs (showcased in this post) and gave it a quick spin with Quake 1. Shared the "achievement" in the same thread next to mpe's.
Here is the full story.
I applied @feipoa's mod to the CPU. With few small improvements.
Didn't cut the pin, but only unsoldered and bent it upwards, making sure it does not touch the heatsink.
Inserted a piece of insulator to form a bed for one of the diode legs.
Precisely bent the diode wires to fit exactly where i wanted them to be.
Used long and tin soldering hot end to ... well ... solder them in place, with minimal collateral damage.
Replaced the original fan with bigger one, otherwise the CPU can hang when running heavy computation tasks.
The original fan seems to be fine for DOS gaming and casual computing.
Basically, made sure that the mod is easily reversible.
Here is the result:
Tested with the next motherboards:
Hope i can obtain one day a PC-Chips M918i, looks like a good performer.
VLI and 8433UUD worked best.
PB4 and LS were flaky, but i think that can be improved on. Now when i am back at it, will spend some time over the weekend and if successful - will update the post.
PVI was disappointing - barely worked before, i know it won't work now too. Not going to bother with it.
Here is how Asus VLI looked like in speedsys with all settings on max:
L1 cache runs in WT mode. It can be switched to WB, but the system becomes unstable.
I remember running it stably in WB before, but not anymore. At least i have this memory of Quake 1 busting 27.1 fps, but now i am starting to doubt myself. Maybe it was WT really, but the Quake version was 1.08 instead of the 1.06 that comes with Phil's benchmarks package.
1.08 has slightly better performance - that was probably it.
Also, swapped the trusted DRC2278E EIDE controller with Holtek.
The Holtek one had issues with write access to SanDisk Ultra CF cards. Switched to Transcend 133 - works like a charm. Delivers 1Mb/s more than DTC.
The video below shows the DTC one. The change was made after that.
And how 8433UUD looked like:
L1 cache in WB mode !
L2 cache is disabled.
When enabled, the system becomes unstable, unless L2 Cache Wait States is set to 3-1-1-1, which turns-out to be actually slower than no cache. Curious.
Verified this very carefully, to make sure i am not missing something here.
Also, once L2 cache is off - the system is just rock stable - ticks fast and well.
All other settings are on max, except IBC DEVSEL# DECODING which had to be set to "medium" (down from "fast"), otherwise IDE controller does not recognize the CF card.
The standard set of benchmarks with VLI + Ark1000VL and 8433UUD + Matrox Millennium and POD100 / 486DX5-160 CPUs, hope to succeed to stabilize Abit PB4 and LuckyStar-486EF, so they can be included as well for better sampling:
Anything that requires floating point math benefits from the POD100 CPU, otherwise AMD DX5 clocked at 160MHz comes ahead.
8433UUD + Matrox is quite ahead in Windows, obviously. Matrox cards are way faster in Win GUI than anything else released in the mid 90ies.
The system lags behind in DOS. Tested with Ark2000PV and S3 Trio64V2/DX in addition to the MGA. Ark was slightly better in Wolf3D - like +0.3 fps and did quite better in PC Player Benchmark (+11 fps @640x480), but was slower in Windows and showed same perf in Quake, Doom, etc. S3 Trio was similar. So showing the MGA numbers here.
Notice that Asus VLI failed the LightWave3D test. This test is extremely sensitive to system instabilities. It catches many things that are often missed by the "traditional" ones. The issue in this case is related to the L1 cache. If disabled - all is well, but then performance is terrible.
For now, i am willing to let it slide, since everything else seems to function perfectly fine and also - this is predominantly a DOS machine, but i am very strict on myself when it comes to stability, so i will be coming back to it. Things must be completely stable - one way or another. 😀
Quake 1 running on VLI + POD100 + Ark1000VL = 25.7 fps. Pretty good.
And here is a quick Q1-PCPBench-3DS-SpeedSys roll with the VLI board:
https://www.petershipkov.com/temp/retro_pc_im … 2gx4_pod100.mp4
While i prefer non-hybrid systems, POD100 will probably stay in the VLI based PC, for extra flavor. Seems to be a good fit.
But if the other two boards that remain to be properly examined outdo it, they may end-up in the PC case instead.
Decided to modify a second P24T to see if the processor i used until now is faulty in some way, given the instability with LW3D on VLI.
All aspects matched between the two CPUs, so the issue seems to be on the motherboard side.
Btw, this mod turned pretty tidy. Satisfied with the result.
Finally - both P24T CPUs didn't boot at 50MHz base frequency in any of the motherboards.