Disruptor wrote on 2023-02-09, 17:40:

Fascinating

Have you tried 40 MHz with LBD# Sampling point of T2 too?

yes, see pic nr. 1 above

Disruptor wrote on 2023-02-06, 19:28:

Breaking News: Now benching 1st time with a bugfixed 3D Mark 99 Max Prior to this test 3D Mark 99 Max raised wrong values on a s […]
Show full quote

Breaking News:
Now benching 1st time with a bugfixed 3D Mark 99 Max
Prior to this test 3D Mark 99 Max raised wrong values on a system without RDTSC command (486, 5x86).
Re: 3dmark99 MegaThread

Great job, Disruptor and mkarcher!

Here is the score of the patched 3DMARK99MAX: 393 3DMarks; 438 CPU 3DMarks.
Because of only 32 MB RAM, a swapping was necessary, so I had to fix the swapfile to a constant size to reach the best possible performance.

I must correct myself about my glquake-scores from the last few days-
all of them are at 640x480 dpi (best score was 26,6 FPS).

Today I started the test once again at 800x600 dpi - the best score is 26,0.

I managed to start the full speedsys-test, too.
This was only possible after re-starting of Win98SE in MS-DOS-mode.
The CPU-score is 66,94 now.

I also did the Quake-II-test, timedemo 1 at 800x600 dpi and no 8-bit (s_initsound 0; snd_restart) - the score is 11,1 FPS

Other results in DOS are:

-> Doom: 65,34 (1143 realtics; no sound)
-> 3d bench 1.0c: 98,5
-> PCP-bench (vga-mode): 23,9

gonzo wrote on 2023-02-10, 15:46:

Great job, Disruptor and mkarcher!

Here is the score of the patched 3DMARK99MAX: 393 3DMarks; 438 CPU 3DMarks.
Because of only 32 MB RAM, a swapping was necessary, so I had to fix the swapfile to a constant size to reach the best possible performance.

It's not my honour. It's mkarcher's!

Well, my 447 CPU Marks on my 486/160 seem to be more than your 438 on your 486/180. However, I use 256 MB FPM with 1024 kB L2 WT Cache.
But it seems GeForce 5200 PCI and Matrox G450 PCI seem to have no chance against your Voodoo 3.

Disruptor wrote on 2023-02-10, 17:07:

It's not my honour. It's mkarcher's! […]
Show full quote

gonzo wrote on 2023-02-10, 15:46:

Great job, Disruptor and mkarcher!

Here is the score of the patched 3DMARK99MAX: 393 3DMarks; 438 CPU 3DMarks.
Because of only 32 MB RAM, a swapping was necessary, so I had to fix the swapfile to a constant size to reach the best possible performance.

It's not my honour. It's mkarcher's!

Well, my 447 CPU Marks on my 486/160 seem to be more than your 438 on your 486/180. However, I use 256 MB FPM with 1024 kB L2 WT Cache.
But it seems GeForce 5200 PCI and Matrox G450 PCI seem to have no chance against your Voodoo 3.

Sadly, my L2 cache tag bits runs at FSB 60 MHz at 8bits only (7bits would be much better, but it crashes) - this is the main reason for the average results of the 180 MHz, together with the poor RAM-throughput in speedsys...
But, anyway, it is not too bad for a fast 486-system, I think.

Maybe you should try 128 MB of RAM in WB-mode, there should be a better CPU-score, or even a better 3D score, too.

Why don't you try to run your L2 in WT?

Disruptor wrote on 2023-02-10, 18:05:

Why don't you try to run your L2 in WT?

This was a very good idea, Disruptor, thank you 😀

This is one of the strangest boards I have: this WT-option changes many things,
now the system becomes faster.

Now, 64 MB RAM are possible. I had to de-crease the DRAM-speed at "normal" (regardless of 32 or 64 MB, regardless of 256 KB or 512 KB L2, and regardless of Samsung- or non-Samsung RAM-chips used).

To make it more funny, I use Samsung chips with 50 ns (but, this seems not to have any effects, as there are no faster timings compared to the 60 ns chips possible). BTW, this is the first of my 72pin-RAM-boards working fine with 50ns chips .

The DX5-133 works at 180 MHz and 5,0 Volt still rock-stable 😀

Sadly, the Voodoo 3 does not work properly with the T2-BIOS-settings any more, so it works now at T3.

But, as good known about 486-systems, in reality you can not have all things at 100 % at the same time...

Anyway, now the scores are:

- speedsys: 101,49 MB/s RAM-throughput and 67,47 cpu-score

- 27,3 FPS in glquake at 800x600

- in Quake II at 800x600: 11,8 FPS with snd_restart/s_initsound 0 and 15,0 FPS with additional gl_flashblend 1/cl_gun 0/cl_particles 0/gl_dynamic 0

- 3DMARK99MAX: 427 3DMarks and 468 CPU 3DMarks

- Sandra98: 39 MB/S RAM-throughput

- doom (no sound): 66,56 FPS (1122 realtics)

- 3dbench 1.0c: 91,1

- PCP-Bench (vga-mode): 25,5

more of them

gonzo wrote on 2023-02-11, 16:57:

Now, 64 MB RAM are possible.

I'm just a bit confused about how your performance gain in 3DMark 99 Max is influenced by the increased amount of DRAM.

Disruptor wrote on 2023-02-11, 18:15:

gonzo wrote on 2023-02-11, 16:57:

Now, 64 MB RAM are possible.

I'm just a bit confused about how your performance gain in 3DMark 99 Max is influenced by the increased amount of DRAM.

I think it is more caused by the BIOS-settings.
If you don't have a Voodoo, you can use a Geforce 2 MX (200/400).
Die Geforce FX does not have good/fast drivers for older DOS-systems, as far as I know (so they are faster than the 2MX only "on paper").
Oh, and I am not sure, if a Geforce FX does start on any 486-board, due to the PCI-version it needs (are they any informations about this combination?).

Very interesting.
I have zida 4dps revision 2.1 and Tim’s pci400-4 which are kind of the same pcb.
Neither one of them handles well anything about 160mhz.
They even have trouble with 4x40mhz.
Can you post a good picture of your board how it is setup for 3x60, also of the memory stick ?

—-

Looking at your setup - if you try to find better quality l2 cache chips that will allow you to furter optimize the current 3-2-3 timings which is what holds you back.
Do you have some spare chips to try ?

pshipkov wrote on 2023-02-12, 05:48:

Very interesting. I have zida 4dps revision 2.1 and Tim’s pci400-4 which are kind of the same pcb. Neither one of them handles w […]
Show full quote

Very interesting.
I have zida 4dps revision 2.1 and Tim’s pci400-4 which are kind of the same pcb.
Neither one of them handles well anything about 160mhz.
They even have trouble with 4x40mhz.
Can you post a good picture of your board how it is setup for 3x60, also of the memory stick ?

—-

Looking at your setup - if you try to find better quality l2 cache chips that will allow you to furter optimize the current 3-2-3 timings which is what holds you back.
Do you have some spare chips to try ?

Yes, of course, phipkov, I am glad to help 😀

Sadly, I do not have faster L2-modules...

BTW, thank you for asking about the RAM-pictures: I realised, that at one side of my 50 ns 64-MB-stick, they are 60 ns chips (in fact, this module was running in 60 ns)!! That's why I thought, they are no better RAM-settings between 50 and 60 ns on this board possible - I was wrong!

I found another module with 50 ns chips only, even a 32-MB-one.
So it is possible to reach better RAM-timings 😉

I found out, that the better scores in all tests are for DRAM-values in the BIOS:

DRAM RAS to CAS Delay: 2
DRAM Write Cycle: 0
DRAM Write CAS Pulse: 1
DRAM CAS Precharge Time: 1
DRAM RAS to MA Delay: 2
DRAM speed: fast

(note: there is also possible a DRAM-speed of "fastest", but the other RAM-parameters must be de-creased, so the resulting RAM-bandwith and RAM-throughput are lower, and this, sadly, affects all scores of all tests, too - so I do not use this "fastest" option)

So, at "fast" RAM-settings, the 3DMARk-scores rises in both CPU (483) and 3D (444) 😀
The AMD 486 seems to be very sensitive regarding the RAM-speed/RAM-parameters.

Once again: at those settings for RAM and CPU, the system is absolutely stable and quick to handle in Windows (no one failure using it or during the tests).

BTW, it seems not to have a big effect for 3DMARK99MAX, if 32 or 64 MB RAM are used (of course, on my Windows-installation they are no many active programs in background, so my RAM is not "overloaded"...)

Anyway, now the scores are:

- speedsys: 130,51 MB/s RAM-bandwidth and 67,48 cpu-score

- 27,3 FPS in glquake at 800x600

- in Quake II at 800x600: 11,9 FPS with snd_restart/s_initsound 0, max. texture quality; and 12,0 FPS at lowest (very ugly) texture quality

- Sandra 98: 49 MB/S RAM-throughput

- doom (no sound): 67,71 FPS (1103 realtics)

- 3dbench 1.0c: 96,9

- PCP-Bench (vga-mode): 26,9

- Quake 1 (DOS): 19,1 FPS

About the board-setting for 60 MHz FSB, I found them here on vogons shown by feipoa here: Zida 4DPS (Tomatoboard) Socket 4 SiS PS/2 mouse not working

My revision 1.0 does not have a PS/2-header.
Also, I was not able to get an USB-mouse working together with an USB-1.1-PCI-controller (regardless if Opti 82C861 or CMD 0670B-400)- the controller can be installed correctly, but no mouse is found, even a powered USB-distributor is connected to the USB-controller first.

more of them

and more

gonzo wrote on 2023-02-12, 15:33:

64 MB, L2 Cache Tag Bits 7 bits

You probably have had a failure in your 64 MB L2 cache configuration. --> just half of your DRAM had been cached in that test run.
You need to use 8 Tag bits with Write Thru.

Disruptor wrote on 2023-02-12, 16:56:

gonzo wrote on 2023-02-12, 15:33:

64 MB, L2 Cache Tag Bits 7 bits

You probably have had a failure in your 64 MB L2 cache configuration. --> just half of your DRAM had been cached in that test run.
You need to use 8 Tag bits with Write Thru.

That's interesting.
Is this specific for this SIS-chipset only, or is this so generally?
Can you maybe do a list (a compilation) of possible L2-BIOS-configurations in accordance to the total DRAM-ammount used (most interesting for 16/32/64/128 MB)?
This could be very helpfull.

gonzo wrote on 2023-02-12, 18:54:

Is this specific for this SIS-chipset only, or is this so generally?

No, it is a general rule since modern 486 chipsets do not have a dedicated dirty tag ram.
Write through --> 8 bits
Write back --> 7 + 1 bit

But I'm still confused why your board does always dirty write back. This should only happen in Write back with 8 bit configuration. But write back with always dirty is seldom a good cache policy.

For 32 bit systems, like 386, 486:

l2 cache SRAM / DRAM
486 write through
1024 kB / 256 MB
512 kB / 128 MB
256 kB / 64 MB
128 kB / 32 MB
64 kB / 16 MB
486 write back
1024 kB / 128 MB
512 kB / 64 MB
256 kB / 32 MB
128 kB / 16 MB
64 kB / 8 MB

For Pentium systems, double the amount of RAM

Limit RAM to 32Mb and try WB L2 cache to see how it goes.

Thanks for the details.
Good to see that this motherboard is capable running well.
Will give the one here a good push at some point soon.
Past attempts were unsuccessful.

Will link to your posts from another thread where i have bunch of posts about 486 hardware.

I've not had good luck with the 4DPS for speed and stability.

Once you think you've found a stable configuration, try installing Windows 2000 to see if it finishes thru to completion without errors. Letting GLQuake run in loop for 2 hours is another telling, yet simple test.

Disruptor wrote on 2023-02-12, 20:18:

gonzo wrote on 2023-02-12, 18:54:

Is this specific for this SIS-chipset only, or is this so generally?

But I'm still confused why your board does always dirty write back.

Disruptor, I think, you are right: the board seems to use always a dirty bit.

Choosing the WB-setting in BIOS, I was able to start only the first time into Windows (as I wrote some days ago, the system is unstable in combination of WB and 7bits).
At this single time in Windows, the Sandra98- and the speedsys-results were very close/equal to the WT-settings, so the dirty bit seems to be always active, regardless of WB or WT in BIOS. Also, I tested 3DMARK99MAX at WB - the score was
equal to WT (443 3D-points), too.

Back to DOS (re-start in DOS-mode from Win98SE), speedsys shows me very similar scores like in WT-mode.
I also did a CTCM- and a CACHECHK-test.

After the next reset ("warm re-start") few seconds later of the system, THE PROBLEMS STARTED immediately:

- no one more booting into Windows at WB
- back to WT+7 bits, Windows started only one time properly; the CTCM-test for WT in DOS was done in one or two seconds only, and the scores in the last three rows were equal!

Now I am using WT+8bits again, as this seems to be the only really stable combination for my system, and as there is no performance-difference to WT+7bits or WB+7bits.
The last three rows in CTCM remains equal...Anyway, all tests in DOS/Windows are successfull.

My question is: why is the board "really stable" at WT+8bits only, although there is no performance-difference to WT+7bits?

I see at least 4 possibilities about the failures:
- defective L2-modules (BTW, can defective L2-cache force a board to work always at dirty-bit-mode?)
- the used L2-modules have problems at FSB 60 MHz (I don't think so, as the system is stable for a longer time during the last few days, even at WT+8bits only)
- strange board
- a BIOS-bug (the 500A can be found here: Zida 4DPS (Tomatoboard) Socket 4 SiS PS/2 mouse not working)

Or (to be more pessimistic), the worst case - maybe a combination of them...

For sure, I used another power-supply, too (as this can be a reason for an unstable system, too)...

BTW, is there any dependency between the L2-modes in BIOS, and the T2/T3-settings of the Voodoo 3?

the last one picture (CTCM for WT+8bits)