Stage 2.
Latest beta BIOS from Asus P5A-B was flashed. UDMA66 works. But, I can not find the settings on 8/10 bit TAG SRAM mode. Asus BIOS should be tweaked for this. 128Mb & 256Mb we’re tested. 128 works faster. Need further tweaking 😬

Got 3200 points.

Mentat-vvo wrote on 2020-05-22, 08:31:

Latest beta BIOS from Asus P5A-B was flashed. UDMA66 works. But, I can not find the settings on 8/10 bit TAG SRAM mode. Asus BIOS should be tweaked for this. 128Mb & 256Mb we’re tested. 128 works faster. Need further tweaking 😬

I guess you know this, however my suggestion is try to unhide the "8/10 bit TAG SRAM mode" option with modbin (e.g. ver. 4.50.82). In the Gigabyte GA-5AX bios, it's just hidden but I have no idea about the Asus bios. If it's not present there you will probably need a patch.

First of, I want to say thanks for Mentat and Lotosdrache for your efforts and time devoted to this!!
Everything we do here is undocumented and unheard of!
So dont be shy to post photos 😁 Mentat I would love to see a photo of the mentioned 1M/512k is this text under the Sram? And what is the original rev of your P5AB?
Yes the information mentioned in the ALI 1541/1542 datasheets explains that you can not use 1Mb cache without extra bits from external tag ram. However ASUS P5A Rev 1.06 does exist with 1Mb cache!
How they made it work I do not know 🙁

I am still trying and failing everyday with the ATI card. one thing I noticed thus far is that you can not run the system at 32bit colour. PC will restart itself(only with seems to be newer ATI drivers)
Cat driver 3.2 is the fastest for my testing thus far but it is not close to GF3 and also not close to the results I got in 2016 using GA5AX and ATI 9250.
I am thinking that with ATI you must use Win2K SP4

Its super awesome to see the amount of ram Mentat was able to POST with 😁 😁

Here are my conclusions of my first benchmark run:

Lotosdrache wrote on 2020-05-21, 09:48:

EDIT 1: I finished benchmark run with original Gigabyte BIOS F4 (see table above). I did most of the runs only once and that wi […]
Show full quote

EDIT 1:
I finished benchmark run with original Gigabyte BIOS F4 (see table above). I did most of the runs only once and that with 512 MiB RAM and DirectX 8.1b twice. However, I think I can make/confirm following remarks/statements:
1. The result range is considerably high. My estimation for a given configuration is: average +/- 50 points. Therefore it's hardly possible to declare a winner.
2. The differences between both DirectX versions are small (128 MiB / 512 MiB):
- DirectX 7.0a: 2923.2 / 2465.2
- DirectX 8.1b: 2898.4 / 2470.1
- Δ..................... -24.8 / 4.9
3. The differences between all 5 drivers are small (128 MiB / 512 MiB):
- M 98SE: 2920 / 2443.7
- ALi 1.72: 2951 / 2457.7
- ALi 1.82: 2916 / 2470
- ALi 1.90: 2890 / 2479.3
- ALi 2.13: 2877 / 2491.7
4. 128 MiB RAM are much faster than 512 MiB RAM (256 MiB RAM gave similar results as 512 MiB):
- 2910.8 <-> 2468,48 / Δ -442,32
5. K6-III+/400ATZ@550 MHz (5.5*100 MHz) is much faster than K6-2+/570ACZ@600 MHz (6*100 MHz) albeit a little bit more unstable (128 MiB / 256 MiB):
- K6- 2+: 2837 / 2475
- K6-III+: 3441 / 3257
- Δ............ 604 / 782
Therefore, more on-die cache is more helpful than higher clock.
6. With K6-2+/570ACZ@600 MHz on GA-5AX, I hardly achieve 3000 points with FireGL 8800 and standard chipset configuration 🙁

To be continued...

Try to test modified BIOS asap 😎

I finished first round of testing my modified BIOS on Gigabyte GA-5AX:

Lotosdrache wrote on 2020-05-21, 09:48:

EDIT 2: I finished testing of my BIOS patch: 1. Differences between both DirectX versions are small again (128 MiB / 512 MiB): - […]
Show full quote

EDIT 2:
I finished testing of my BIOS patch:
1. Differences between both DirectX versions are small again (128 MiB / 512 MiB):
- DirectX 7.0a: 2857.6 / 2848.2
- DirectX 8.1b: 2875.6 / 2888.6
- Δ...........................18 / 40.4
3. The differences between all 5 drivers are small again (128 MiB / 512 MiB):
- M 98SE: 2857 / 2856.5
- ALi 1.72: 2847.5 / 2839
- ALi 1.82: 2881 / 2894.5
- ALi 1.90: 2826.5 / 2891
- ALi 2.13: 2921 / 2861
4. 128 MiB RAM are no longer faster than 512 MiB RAM:
- 2866.6 <-> 2868,4
- Δ 1.8
😀
From this I can conclude that
- the patch works and the board is now able to cache 512 MiB of memory with its internal tag RAM 😀
- decreased results with original BIOS F4 are possibly due to smaller cacheable area
5. Onboard L2 cache (= L3 cache with K6-2+) still impacts performance even with K6-2+ and its internal L2 cache
When switching onboard cache (=L3 cache with K6-2+) off, I only got 2527 points.
6. Upgrading RAM to 1 GiB decreased performance again:
- 2474 -> Here I am again out of the cacheable area with MS Win98SE original files.
Interestingly, using Rudolph Loew's patchmem files for systems with more than 512 MiB RAM kept the result high:
- 2837 -> Either the patch limits the system to the lower cacheable 512 MiB, or it reprograms Win98SE to fill the cacheable area first. I don't know.

All in all:
- DirectX 7.0a seems to be less stable on this system than DirectX 8.1b. At least, I got more abnormal terminations with version 7.0a than with 8.1b.
- With the new BIOS, the system seems to be a little bit slower than with the older one albeit of faster RAM timings. At least, I never reached 3000+ points as I did once before with 128 MiB RAM.
Therefore, I have to check whether SDRAM Enhanced Page Mode is really so enhanced. I tested 4 CPUCLKs here. Other settings are 8, 12, or 16 CPUCLKs. However, I never saw a board where this is enabled. Maybe I will disable it again. For this tests I'm going to use a Matrox G400 Max...

Next round will start next week...

That is awesome data you have added to this thread! 😁
And the 1Gb ram part is epic 😁

Here is driver 7.76 and 7.58 😀 It has already been modified to detect Ti cards.
7.76 can work on Geforce 2/3 Ti
7.58 Geforce 2 Ti

Skalabala wrote on 2020-05-28, 11:24:

First of, I want to say thanks for Mentat and Lotosdrache for your efforts and time devoted to this!! Everything we do here is u […]
Show full quote

First of, I want to say thanks for Mentat and Lotosdrache for your efforts and time devoted to this!!
Everything we do here is undocumented and unheard of!
So dont be shy to post photos 😁 Mentat I would love to see a photo of the mentioned 1M/512k is this text under the Sram? And what is the original rev of your P5AB?
Yes the information mentioned in the ALI 1541/1542 datasheets explains that you can not use 1Mb cache without extra bits from external tag ram. However ASUS P5A Rev 1.06 does exist with 1Mb cache!
How they made it work I do not know 🙁

There're SRAM chips 512KB (64K x 64 UT6164C64) and 1024KB (128K x 64 UT61128C64). To upgrade external cache up to 1MB there should be wires on a mainboard's PCB to connect an additional address line of SRAM 1MB chip to a tag SRAM, a CPU and a NorthBridge.

Roman555 wrote on 2020-06-07, 17:41:

There're SRAM chips 512KB (64K x 64 UT6164C64) and 1024KB (128K x 64 UT61128C64). To upgrade external cache up to 1MB there […]
Show full quote

Skalabala wrote on 2020-05-28, 11:24:

First of, I want to say thanks for Mentat and Lotosdrache for your efforts and time devoted to this!! Everything we do here is u […]
Show full quote

First of, I want to say thanks for Mentat and Lotosdrache for your efforts and time devoted to this!!
Everything we do here is undocumented and unheard of!
So dont be shy to post photos 😁 Mentat I would love to see a photo of the mentioned 1M/512k is this text under the Sram? And what is the original rev of your P5AB?
Yes the information mentioned in the ALI 1541/1542 datasheets explains that you can not use 1Mb cache without extra bits from external tag ram. However ASUS P5A Rev 1.06 does exist with 1Mb cache!
How they made it work I do not know 🙁

There're SRAM chips 512KB (64K x 64 UT6164C64) and 1024KB (128K x 64 UT61128C64). To upgrade external cache up to 1MB there should be wires on a mainboard's PCB to connect an additional address line of SRAM 1MB chip to a tag SRAM, a CPU and a NorthBridge.

Ah thats awesome info! Thanks a lot 😀
I however think that the LP61 64128F-4 will be a better chip, I suspect that the 1Mb chips will be unstable.

Skalabala wrote on 2020-06-07, 19:24:

Ah thats awesome info! Thanks a lot 😀

It's my pleasure 😀

Also it might be necessary to change hardware setup of cache size or internal/external tag.
This data is missing in the datasheet of M1541-M1542. But you can get an idea from the datasheet of M1531 and the picture.
I've just found the hardware setup on different page.

Roman555 wrote on 2020-06-07, 20:28:

Also it might be necessary to change hardware setup of cache size or internal/external tag. This data is missing in the datashe […]
Show full quote

Also it might be necessary to change hardware setup of cache size or internal/external tag.
This data is missing in the datasheet of M1541-M1542. But you can get an idea from the datasheet of M1531 and the picture.
I've just found the hardware setup on different page.

As stated in the data sheet of M1541, section 4.2, register index 40 and 41, this can also be done in software as I have shown above via BIOS patch. I found this way to be safer than soldering because for me it was difficult to identify the correct trace and I didn't want to change original board layout to much. 😉

Roman555 wrote on 2020-06-07, 17:41:

There're SRAM chips 512KB (64K x 64 UT6164C64) and 1024KB (128K x 64 UT61128C64). To upgrade external cache up to 1MB there […]
Show full quote

There're SRAM chips 512KB (64K x 64 UT6164C64) and 1024KB (128K x 64 UT61128C64). To upgrade external cache up to 1MB there should be wires on a mainboard's PCB to connect an additional address line of SRAM 1MB chip to a tag SRAM, a CPU and a NorthBridge.

If I understand you right here, it is necessary to modify the resistor network when replacing the 64K64 L2 data chip by a 64K128 one.
Is this necessary just to identify the correct size of the chip? In this case I think, it can also be done by reprogramming register 41 bits 5 and 3-2 via BIOS patch.
Or is it necessary that the hole size of 1024 kiB L2 cache can be used? Then of course, this cannot be done in software. 😉

That's how my Gigabyte GA-5Ax and Jetway J542C look like:

I like to replace both L2 data chips with one of this:

How do I have to modify/populate the boards for this if software solution is not practicable?
In the case of GA-5AX, does it suffice to remove resistor R58 right above Tag RAM and place it at R59?
What's about J542C? It hasn't even got a Tag RAM.

I think Rev G does not need a modification.(No external tag)
To my knowledge the only boards to have 1Mb Cache ALi is Rev G on ASUS P5A.
There is no other proof of the existence thus far. Only P5A Rev G

Lotosdrache wrote on 2020-06-08, 09:19:

How do I have to modify/populate the boards for this if software solution is not practicable?
In the case of GA-5AX, does it suffice to remove resistor R58 right above Tag RAM and place it at R59?
What's about J542C? It hasn't even got a Tag RAM.

Hi, Lotosdrache
I've never done upgrade of cache. So there could be underwater rocks we don't know about (I mean both the schematics and BIOS code).
1) First of all I want to say that upgrade the cache up to 1MB won't increase performance significatly.
According to these funny marketing diagrams (for previous chipset - ali alladin iv) the increase was by 1% - 5.5%
https://web.archive.org/web/19990429072908/ht … .com/mb575.html
The external tag SRAM is usually 8bit so the cacheble area doesn't increase at all.
2) GA-5AX -
Try to understand if pin 52 of LP61G6464 goes somewere to unpopulated pads nearby (R46, R48, R239, R59 or others). The goal is to find a place for a zero Om resistor that does the same as the R171 of the schematic manual (connects extraline of 1MB cache chip to other chips). So another pin of such resistor should be connected to R59 of GA-5AX. If pin 52 of LP61G6464 connects nowhere - that's the end of research.
If R59 is connected to the address line HA19 and if R58 connects pin 1 of tag SRAM to GND rail then remove the resistor from R58 and add it at R59 place. (R170 and R169 of the schematic diagram accordingly).
3) J542C -
Because the revision of ali M1542 allows omitting the external tag SRAM. The algorithm of research the same. But according to chipset datasheet M1541-M1542 the external tag SRAM is obligatory. So you might end up adding to the mainboard the tag chip (and some components that are necessary for its correct work).
4) The straps (hardware setup) might be necessary. Somehow the motherboard should get L2 size right. BIOS may just get size of cache reading registers that the chipset has filled from hardware setup data.

P.S. Of course everyone who wants to solder a big chip in 128 pin PQFP package should have proper soldering equipement in hand (not only a soldering iron but at least a preheater and a hot air rework station).

P.P.S. It's better to compare datasheet of installed SRAM and new one but they're not easy to find.

The reason I think to upgrade to the 1Mb cache is for the possibility to have more breathing room for the CPU.
Because the CPU cache is really big on K6-3+ 256Kb running at 600Mhz+ Might get chocked by the slow 512Kb.
But it could also not give a big performance boost because of more data to flush.
For now I think 1Mb is a good idea. It might be more stable(needs to be flushed twice less than 512kb) but it also might be less stable(the architecture of the chip itself)

P.S. My world record motherboard is not happy, all these years on 4.0V VIO has taken its toll. I am having a lot of instability problems 🙁
I could run 3dmark at 633Mhz with only 2.0V!! Now iam lucky if I get it done at 2.3V 🙁

Hello again. I’ve recently got my Epox 5MVP3G2 , also known as NMC-5VMXe. It rocks! But, I have “slow AGP” issue. Gonna try different 4in1.

My 3DMark2000’s score is 2600 with this mobo. K6-3+ 620 124*5 with whole 1Mb of cache enabled. It also works @133, but w/o cache.

WOW ! Mentat, that is a good pie time!
Are you saying that you can run 124Mhz on your single 1024kb cache chip? That is some awesome info!

Skalabala wrote on 2020-06-30, 12:35:

WOW ! Mentat, that is a good pie time!
Are you saying that you can run 124Mhz on your single 1024kb cache chip? That is some awesome info!

Yep. It also runs at 133 MHz WITH cache enabled. 3DMark passes Game1 &2 @ 633, but SuperPi does not ) Maybe, that was memory issue. Need Clean install of Win2k, where system gets +20 mb/s memory boost over Win98.

Very nice. The highest I was ever able to get was 120x5.5 on a P5A with a K6-2 550.. but that was back when that stuff was brand new.

Got ~4600 points with the right driver and GF3Ti200.
And about 100 points more with Quadro2Pro.
Now, I’m learning how to apply 2-2-2-5 delays to RAM and other tweaks via WPCREDIT.

UPDATE: I've managed to set 2-2-2-5 delays to SDRAM:

Thats the spirit 😁 Good work!
Care to share photo of your CPU? And is that score with 112x5.5 or 124x5 or 133x4.5?
My best score with MVP3-M is 4800+ with 112x5.5

Try this to tweak maybe.
https://drive.google.com/file/d/1VI9ot2yN9v1h … iew?usp=sharing