Great result with the Thunderbird CPU!

cyclone3d wrote on 2026-07-27, 08:04:

Gave up on getting the nf7 2.0!for now.

Setting up the EPOX EP-8VTAI (later renamed to 8KRAI Pro). The chipset is AMD KT880. BIOS is weird and I will probably see if I can mod it to at least unlock hidden features.

The Socket A boards with VIA chipsets also use romSIP tables in the BIOS like the NF2 boards. So in theory, you can swap to faster SIPs. I guess it depends what table the BIOS use.
The tables are in the main BIOS at offset 74000h.

I tested different SIP tables with my ASUS KT600 board in the past. Faster SIPs reduce OC FSB speed, though.
https://www.hardwareluxx.de/community/threads … 4#post-31028709 (german)

_digitalbath wrote on 2026-07-28, 14:14:

Great result with the Thunderbird CPU! […]
Show full quote

Great result with the Thunderbird CPU!

cyclone3d wrote on 2026-07-27, 08:04:

Gave up on getting the nf7 2.0!for now.

Setting up the EPOX EP-8VTAI (later renamed to 8KRAI Pro). The chipset is AMD KT880. BIOS is weird and I will probably see if I can mod it to at least unlock hidden features.

The Socket A boards with VIA chipsets also use romSIP tables in the BIOS like the NF2 boards. So in theory, you can swap to faster SIPs. I guess it depends what table the BIOS use.
The tables are in the main BIOS at offset 74000h.

I tested different SIP tables with my ASUS KT600 board in the past. Faster SIPs reduce OC FSB speed, though.
https://www.hardwareluxx.de/community/threads … 4#post-31028709 (german)

I think what I need is a SIP that will force the 200Mhz profile on lower stock fsb CPUs. This board seems very dodgy with Thunderbird above about 170Mhz fsb. Highest I have been able to attain is a score of 13,000 but I know there is way more there.

I did switch to a 7950GT and found that it doesn't have the huge pauses / massive slowdowns in the first couple benchmarks like the HD3850 does. The same exact slowdowns were happening on the nforce 2 ultra boards. I'm not sure if that is normal for the HD3850 or if it was because that card seemed flakey.

I haven't completed a bench run with the 7950GT yet but the overall fps are expectedly lower.

In my opinion, the SIP tables I saw in the BIOS were not crucial. I haven't tested any VIA SIP tables with a Thunderbird CPU yet.You might indeed need to replace the first table with a "slower" one.I have no clue whether the tables are selected based on the CPU's default FSB speed or just the current FSB clock. Looking at NForce2 (NF2) BIOSes, most of them select tables by CPU type. That being said, it doesn't have to be that way for every chipset and BIOS.
AMI NForce2 BIOSes select the table by the system FSB speed, so I guess that might be the case here as well.

As far as I remember, the BIOS has four tables. Easiest guess: >100MHz, >133MHz, >166MHz, >200MHz.
Changing SIP tables in your BIOS, you just have to copy paste the whole block code with a hex editor. To be sure, just go back to modbin and safe the BIOS under a new name. Modbin does the checksum for you.

Soni would want to swap out the 100 and 133 with either the 166 or 200 for testing?

Easy enough to use my USB flasher if something goes wrong.

cyclone3d wrote on 2026-07-29, 17:34:

Soni would want to swap out the 100 and 133 with either the 166 or 200 for testing?

Easy enough to use my USB flasher if something goes wrong.

The first table is at offset 74000h, others following. The first table is what I posted in a previous post. I would try to replace it with the code from the following three tables. The followings tables should be the "slowest" SIP.

_digitalbath wrote on 2026-07-29, 18:23:

cyclone3d wrote on 2026-07-29, 17:34:

Soni would want to swap out the 100 and 133 with either the 166 or 200 for testing?

Easy enough to use my USB flasher if something goes wrong.

The first table is at offset 74000h, others following. The first table is what I posted in a previous post. I would try to replace it with the code from the following three tables. The followings tables should be the "slowest" SIP.

There are 4 files in that zip ? NM, the _21F is the first one and is included in the zip.

Since the 1333 is a 133fsb CPU, I would think I would want to replace the 1st and second tables. Probably with the slowest table and then if that works well, try the 166 table to see if that lowers the max fsb for the T-Bird.

Not sure what CPU I am going to test after I get the T-Bird at max fsb... I've got a bunch of different mobile, geode NX, etc. Basically just about everything that I would want to test. We could even add different Duron to the list though I only have a couple and nobody else has bothered with them.

Edit:
Actually, only the first SIP in this BIOS is the faster one. Looks like it is probably based on the CPU type detected. So I only need to replace the first one.

Edit2: Actually, trying the first SIP in the second block so it supposedly will apply the 100Mhz SIP, which is supposedly looser timings, in the 133Mhz block. The VIA chipset BIOSes supposedly use the ROMSIP based on the detected stock fsb.

The first one being the only one that is different is suspicious and I am guessing that they did not actually test with T-Bird CPUs at all.

cyclone3d wrote on 2026-08-01, 00:35:

Edit2: Actually, trying the first SIP in the second block so it supposedly will apply the 100Mhz SIP, which is supposedly looser timings, in the 133Mhz block. The VIA chipset BIOSes supposedly use the ROMSIP based on the detected stock fsb.

Usually the lower FSB sips got tighter timings, as it clocks slower. You got more realtime per clock and thus need fewer clocks for the same amount of real time to pass. So i'd expect the 200Mhz FSB ROMSIP to be the slowest of the bunch. Note that going too slow might also cause timings issues and even a no-POST. So for a 133Mhz T-Bird, i'd start with 166Mhz Sips and then 200Mhz.

You could also try the socket wire mod (basically "hard" L12 bridge mod) before modding the bios itself.

Ok, so took a gander at the ASUS A7V880 BIOS and it has a total of 16 ROMSIP tables.

ChatGPT claims it can decode the ROMSIPs based on the documentation for the KT880. Going to try on of the Asus ones that is apparently the correct one for the T-Bird CPUs and has more wait states that will specifically help with the T-Bird CPUs.

cyclone3d wrote on 2026-08-01, 00:35:

There are 4 files in that zip ? NM, the _21F is the first one and is included in the zip.

True, I've copied all variants, that I found in several BIOSes, to the file. You can choose what table you want.
As written, faster SIPs limit the OC.
slowest = D95 < 217 < 21F < 69F =fastest

So, the BIOS has by default the second slowest and following by the slowest SIPs. In my oppinion, they should not limit your overclocking attempts.
I guess selecting faster SIPs for the Thunderbird CPUs may be good as they will limit your FSB anyway at a point.

I used to select a single table version for all tables in the BIOS. This method helps me to be sure the CPU uses the table that I want. 😀

cyclone3d wrote on 2026-08-01, 08:09:

Ok, so took a gander at the ASUS A7V880 BIOS and it has a total of 16 ROMSIP tables.
ChatGPT claims it can decode the ROMSIPs based on the documentation for the KT880. Going to try on of the Asus ones that is apparently the correct one for the T-Bird CPUs and has more wait states that will specifically help with the T-Bird CPUs.

Many VIA boards have a lot of tables. No clue why.
The main difference that the VIA tables over the Nforce table have, they have timings for multiplier 3 and 4. You can use a mobile CPU (or Geode) and lower the multi down to 3.

Well, with T-Birds and the 8VTAI, the D95 SIPs seem to be stable up to about 170Mhz fsb and all the others are stable up to 150Mhz.

Past that and stability tanks hard.

Seems like this board is just pretty horrid for T-Bird.

cyclone3d wrote on 2026-08-01, 08:09:

ChatGPT claims it can decode the ROMSIPs based on the documentation for the KT880.

I have not found a KT880 datasheet in years of searching. I got KT600 and older. AFAIK VIA kept the basic Romsip scheme till KT600, but changed it for 880.

We also got a datasheet for nforce 2, but not Ultra 400.

So if any AI claims it can decode the Romsips, I’d proceed with caution and double check everything.

The KT880 BIOSes, that I saw, used the same SIP tables (all bits are the same) as the other VIA socket A boards (KT266A - KT600). VIA changed all PCI register with the KT800 chipset. Maybe VIA changed something different too?
Imho, AI has no clue about romsips. My questions were all wrong.

cyclone3d wrote on 2026-08-02, 04:30:

Well, with T-Birds and the 8VTAI, the D95 SIPs seem to be stable up to about 170Mhz fsb and all the others are stable up to 150Mhz.

Past that and stability tanks hard.

Seems like this board is just pretty horrid for T-Bird.

strange.

_digitalbath wrote on 2026-07-28, 14:14:

I did switch to a 7950GT and found that it doesn't have the huge pauses / massive slowdowns in the first couple benchmarks like the HD3850 does. The same exact slowdowns were happening on the nforce 2 ultra boards. I'm not sure if that is normal for the HD3850 or if it was because that card seemed flakey.

My HD3850 does the same. I haven't messed around with too many drivers yet, as I have had to modify one and write an ATI initialisation-delay driver to make it boot on my rig repeatedly with all the go-faster bits turned on. My 7800GS has been fine, albeit not as high on peak FPS.

Tzk wrote on 2026-07-05, 07:02:

Please do not use Panasonic FR for the cpu. […]
Show full quote

Please do not use Panasonic FR for the cpu.

Usually Kemet A750 16V 2200uF got a low enough ESR (around 10 to 20mOhms) and will fit great as a replacement for the old Nichocon HM or Rubycon MBZ. Don’t worry that the old caps are 6.3V 3300uF, the lower capacity of the Kemet is not an issue. the resistence (ESR) is the primary factor hat higher switching frequencies. A7N8X switches at about 125kHz, other boards switch even faster.

The Kemet will work on the cpu side for a number of boards, just don’t replace all caps on the whole board with Kemet.

Make sure to also replace the 6.3V 1800uF capacitor below the AGP slot on NF7. That one is dead on most boards and the one which usually keeps the board from POSTing. Also use Kemet there, as it’s very low ESR.

As I plan to recap the A7N8X, what is the main issue with the Panasonic 6.3V 3300uF? And what about these: 3300/6.3PHT(L.E)PANASONIC?

Tzk wrote on 2026-08-02, 08:52:

I have not found a KT880 datasheet in years of searching. I got KT600 and older. AFAIK VIA kept the basic Romsip scheme till KT6 […]
Show full quote

cyclone3d wrote on 2026-08-01, 08:09:

ChatGPT claims it can decode the ROMSIPs based on the documentation for the KT880.

I have not found a KT880 datasheet in years of searching. I got KT600 and older. AFAIK VIA kept the basic Romsip scheme till KT600, but changed it for 880.

We also got a datasheet for nforce 2, but not Ultra 400.

So if any AI claims it can decode the Romsips, I’d proceed with caution and double check everything.

Could I get a copy of the KT600 datasheet?

cyclone3d wrote on 2026-08-03, 00:10:

Could I get a copy of the KT600 datasheet?

It's uploaded on retroweb, just click in "4 chip documents available" and download the PDF.

https://theretroweb.com/chipsets/542

SlowA wrote on 2026-08-02, 16:56:

As I plan to recap the A7N8X, what is the main issue with the Panasonic 6.3V 3300uF? And what about these: 3300/6.3PHT(L.E)PANASONIC?

The ESR of 18mOhms is too high. The stock Nichicon HM or Rubycon MBZ 6.3V 3300uF caps got about 10-13mOhms of ESR.

Excellent. Using the KT600 datasheet and ROMSIP tables from KT400 - KT880 based boards shows that the ROMSIP register layout did not change.

Next up is to compare specific PCI register changes based on different ROMSIPs and then test with single byte changes.

Tzk wrote on 2026-08-02, 08:52:

We also got a datasheet for nforce 2, but not Ultra 400.

My google-fu is quite poor and I have been unable to find this. Where is it? Can you help?

EduBat wrote on 2026-08-03, 14:36:

Tzk wrote on 2026-08-02, 08:52:

We also got a datasheet for nforce 2, but not Ultra 400.

My google-fu is quite poor and I have been unable to find this. Where is it? Can you help?

I would like this as well. The only thing I am finding is reverse engineered data.

The ESR of 18mOhms is too high. The stock Nichicon HM or Rubycon MBZ 6.3V 3300uF caps got about 10-13mOhms of ESR.

Unfortunately only Panasonic is available here, what is the worst outcome if I recap with Panasonic FR?