The picture is from another C3 thread:

Cyrix C3 and 486/AMD K-6 builds in progress

My Ezra 900 is a match of that one. So far I got:

1 x Ezra 800 (133 x 7.5)
1 x Ezra 900 (100 x 9)
1 x Nehemiah 1200 (133 x 9)

The goal is to get:

1) the Ezra 900 to support 12 * 100 and 9 * 133
2) the Nehemiah to support 12/14 * 100

The Ezra 900 is a good starting point, since its native multiplier is 9, so if we can get the FSB at 133 will get to 1200. We know from testing, it doesn't tolerate 12 x 100 at 1.35v, so we might need to bump the voltage first, was thinking of 1.45v. Looking at the Ezra datasheet (Chapter 4, Table 4-2), it seems that lifting the VID1 pin (1->0), will take it to 1.45v. Did any of the cool guys already try something like this ? 😀

If once overvolted is stable at 12 x 100, the next step would be to find a way to change the FSB between 100 and 133 Mhz. Will connect a switch either to one of the bridges (if we can confirm that controls the FSB strap) or something creative to turn pin BSEL1 on (133) and off (100). When it boots at 133, multiplier is already correct (9), when it boots at 100, will then use SetMul 12 to take it to 1200.

The Ezra 800 is a fallback, since the specs say that some speed bins are hard-coded to 100, so in case BSEL1 is locked on the 900 (limiting FSB to 100) and we cannot set FSB via the bridges, at least we know that on the Ezra 800 (which is a native 133) BSEL1 in unlocked, so we can use it to choose between 100 and 133.

For the Nehemiah 1200 the goal is to strap the FSB at 100 and set multiplier to 12. Here as well the datasheet shows BSEL1 decides between FSB 100 and 133. So when pulling the pin to 0, will get 100 x 9 at start, and then with SetMul will take it to 1200 (or 1400, which I found super-stable at 133 * 10.5, will see at 100 x 14).

The reason to switch the FSB to 100 is that Voodoo2 doesn't like 133, so many DOS Glide games don't run with FSB at 133, but are totally happy with a FSB at 100 (at the same total clock, e.g. Glide DOS Test[00-29] runs fine at 100 * 10, but Glide can't find the Voodoo2 at 133 * 7.5, both being 1000 Mhz).

Thanks for your help!

See attached image for how Bus frequency select pins (BSEL) are wired on a similar Ezra.
There are two bridges that configure the signal of the Socket 370 BSEL pins, all marked in Yellow. I verified it with a multimeter.
Image is based on screenshots from "VIA C3 C5B Samuel 2 Datasheet V1.12.pdf"

Fantastic! 😀

Does the bridge on the left match BSEL1 and the one to the right BSEL0 ? Is bridge closed equal to 1 on the pin ?

Do we think other bridges might also control voltage (e.g. VID1) ?

sunmax wrote on 2024-08-18, 16:44:

Fantastic! 😀

Does the bridge on the left match BSEL1 and the one to the right BSEL0 ? Is bridge closed equal to 1 on the pin ?

Do we think other bridges might also control voltage (e.g. VID1) ?

The bridge on the left matches BSEL1 and the one to the right BSEL0. Indeed.
A 133MHz FSB CPU has both bridges cut. So 1 in the table means cut. Penciling both bridges should make it a 66MHz FSB request signal.

Haven't got time for the voltage tracing now...

I often use these CPUs on a slotket adapter (on a Slot 1 Motherboard), which exposes the mentioned configuration pins to jumpers.

Great, thanks!

I'll still need to find a way to boost the voltage, we know that at 1200 (setmul=12 x 100) this CPU locks, so once the bridge is cut at native 9 x 133 (1200) will probably not post.

I was planning to trace continuity with a multimeter between bridges and VID1, when I got side-tracked by some MTRR findings: Ezra appears to be substantially faster than Ezra-T (with WC on graphics memory)!

I started a separate thread for that: Re: VIA C3 Ezra vs Ezra-T: Interesting Behavior of MTRR/FASTVID

While studying the MSRs I ran into 2Ah and its control bits. I see there is one to change BSEL from 100 to 133.

Do you think is possible to extend SetMul to also control the BSEL bit, so we could switch C3 at run-time between FSB 100 and 133, as we currently do for multipliers ?

sunmax wrote on 2024-08-26, 18:32:

I was planning to trace continuity with a multimeter between bridges and VID1, when I got side-tracked by some MTRR findings: Ezra appears to be substantially faster than Ezra-T (with WC on graphics memory)!

I started a separate thread for that: Re: VIA C3 Ezra vs Ezra-T: Interesting Behavior of MTRR/FASTVID

Interesting. I don't know if I have the proper benchmarks of Samuel 1/2 + Ezra to settle this matter.
I did provide them for Ezra-T and Nehemiah in the 3rd post of this topic. It says "LFB write combining enabled". Nehemiah does a lot better there.
So basically your suggesting this strangely slow performance issue with Ezra-T was not so much present in Ezra, and is mostly write combining MTRR related?

sunmax wrote on 2024-08-26, 18:32:

While studying the MSRs I ran into 2Ah and its control bits. I see there is one to change BSEL from 100 to 133.
Do you think is possible to extend SetMul to also control the BSEL bit, so we could switch C3 at run-time between FSB 100 and 133, as we currently do for multipliers ?

I don't think it works that way, because it is not within the functionality of motherboard PLL clock generators.
For example this ICS PLL: https://www.renesas.com/us/en/document/dst/9148-26-datasheet
It is not about any CPU MSR, if you want to change FSB dynamically, without reboot, you need software written to communicate with a PLL I2c through SMBus. But only when the PLL supports such features. See Rayer's SMB http://rayer.g6.cz/programm/programe.htm,

Very interesting. Looking at the Ezra-T numbers in the third post, LFB WC would appear to be fully operative for Quake and PCP, since based on the data gathered, those results are in the range of ~38-40 FPS for Quake1 (vs < 30) and for PCP ~50 FPS (vs ~35), of an Ezra at 1Ghz+ with a good PCI/AGP card and LFB WC on. I see you were using a Voodoo3, that's a stellar performer in those tests, so allow me to re-run the benchmarks with a Voodoo3 to see if she's the reason of the higher FPS in place of actual LFB WC. Do you recall if the Ezra-T tested was a native FSB133 x 7.5 and you bumped the multiplier to 9 with SetMul ? Was it overvolted ? Do we have the part number on the back ? The only Ezra-T I've been able to test is an Ezra-T 10 x 100. Yet others in Vogons reported the same unexpected behavior of Ezra-T on their build with VGA/LFB WC, and just minor speed bump turning WC on.

I also found a GA-6BXC, so will later be able to more closely match the setup in the third post. That will also tell us if depends on the CPU or mobo chipset.

Thanks for the clarification about BSEL.

sunmax wrote on 2024-08-29, 23:49:

Do you recall if the Ezra-T tested was a native FSB133 x 7.5 and you bumped the multiplier to 9 with SetMul ? Was it overvolted ? Do we have the part number on the back ? The only Ezra-T I've been able to test is an Ezra-T 10 x 100. Yet others in Vogons reported the same unexpected behavior of Ezra-T on their build with VGA/LFB WC, and just minor speed bump turning WC on.

Must have been these two CPUs, that were used in those 2014 benchmarks:
C3 Neh. 1000 MHz; 133 FSB; 7.5x; 1,45V; 6.9.3; '03-17; AG60BH0
C3 Ezra-T 1000 MHz; 100 FSB; 10.0x; 1,45V; 6.8.9; '02-42; AG60ABO

I don't remember the Voltage set, but I am not charmed by the idea of overvolting, so normally I avoid doing that.

Excellent. I got the same exact Ezra-T, down to production date 😀

Made some additional testing, also included PCP numbers, with post and tables in the VIA C3 WC thread.

Re: VIA C3 Ezra vs Ezra-T: Interesting Behavior of MTRR/FASTVID

In a nutshell, although I didn't have a chance to setup a GA-6BXC system yet and re-bench with a Voodoo3, based on the results, I'm inclined to think the FPS for Ezra-T in post 3, are tied to the higher baseline performance of the 440BX and Voodoo3, more than MTRR WC being fully operative on VGA and WC.

The Nehemiah is in average 10-20% faster than Ezra clock per clock in DOS VGA & SVGA, when WC is enabled. The only way we can get a 100%+ improvement in PCP SVGA and a 80%+ improvement in Quake SVGA is if we are comparing Nehemiah with WC on to an Ezra (non-T) with WC off, or to an Ezra-T with WC on yet severely underperforming.

If there is a chance (I know it's 10yr later), could you verify these findings on the original system ? Possibly also introducing an Ezra (non-T) and clocking them all at 1000 (unless you have a lucky Ezra non-T stable at 1200).

If we can confirm this, the next phase would be: why doesn't WC produce the same boost in Ezra-T, as it does for Ezra and Nehemiah ?

Thanks a lot!

Hi, thanks for the software, you saved me as the old machine I used with dos+98 died.

I turned an athlon xp 2000+ Thorton into XP-M (L5 mod), as A7V600-X motherboard (bios 1002) with 512mb ddr, sound blaster live player 1024 pci and Nvidia Geforce FX 5200. Works great!
the only problem I found doing tests is that if I try to restore the original multiplier (12.5) without rebooting, it sets it to 11
(win98 + gpuz vintage). It's not serious but I'm wondering if anyone has encountered the same problem and/or it can be solved perhaps by updating the bios.

For information, why can't you lower the multiplier below 4x with K7 CPUs?

thanks!

5a796d wrote on 2024-10-20, 15:44:

Hi, thanks for the software, you saved me as the old machine I used with dos+98 died. […]
Show full quote

Hi, thanks for the software, you saved me as the old machine I used with dos+98 died.

I turned an athlon xp 2000+ Thorton into XP-M (L5 mod), as A7V600-X motherboard (bios 1002) with 512mb ddr, sound blaster live player 1024 pci and Nvidia Geforce FX 5200. Works great!
the only problem I found doing tests is that if I try to restore the original multiplier (12.5) without rebooting, it sets it to 11
(win98 + gpuz vintage). It's not serious but I'm wondering if anyone has encountered the same problem and/or it can be solved perhaps by updating the bios.

For information, why can't you lower the multiplier below 4x with K7 CPUs?

Glad you found SetMul useful 😀

K7 has 32 choices for Multiplier. This is the full use of a 5 bit pattern. All these settings are accounted for in the SetMul table for K7. Basically that table should be the same as the internal table baked into the AMD K7 CPU..
When I look at the table in the source code, I have no reason to think 11x and 12.5x are set wrongly there. Though I currently have no assembled system to test again.
What does it do when you specify 11x? Maybe try https://crystalmark.info/en/download/#CrystalCPUID to compare. The BIOS should not matter for this small issue, it is not like the BIOS is intercepting and changing the multiplier request code send from SetMul to the CPU..

5a796d wrote on 2024-10-20, 15:44:

For information, why can't you lower the multiplier below 4x with K7 CPUs?

K7 Should allow 3x, but not 3.5x

5a796d wrote on 2024-10-20, 15:44:

the only problem I found doing tests is that if I try to restore the original multiplier (12.5) without rebooting, it sets it to 11

This has to do with the L6 bridge on the top of the chip and not with SETMUL.

See here for info... If you find a safe way to cut closed bridges, please let me know, I am also interested in doing this.

I could be wrong and there may be a way of forcing SETMUL to ignore the L6 bridge. @gerwin - can you weigh in on this?

gerwin wrote on 2024-10-21, 22:25:

Glad you found SetMul useful :) […]
Show full quote

5a796d wrote on 2024-10-20, 15:44:

Hi, thanks for the software, you saved me as the old machine I used with dos+98 died. […]
Show full quote

Hi, thanks for the software, you saved me as the old machine I used with dos+98 died.

I turned an athlon xp 2000+ Thorton into XP-M (L5 mod), as A7V600-X motherboard (bios 1002) with 512mb ddr, sound blaster live player 1024 pci and Nvidia Geforce FX 5200. Works great!
the only problem I found doing tests is that if I try to restore the original multiplier (12.5) without rebooting, it sets it to 11
(win98 + gpuz vintage). It's not serious but I'm wondering if anyone has encountered the same problem and/or it can be solved perhaps by updating the bios.

For information, why can't you lower the multiplier below 4x with K7 CPUs?

Glad you found SetMul useful 😀

K7 has 32 choices for Multiplier. This is the full use of a 5 bit pattern. All these settings are accounted for in the SetMul table for K7. Basically that table should be the same as the internal table baked into the AMD K7 CPU..
When I look at the table in the source code, I have no reason to think 11x and 12.5x are set wrongly there. Though I currently have no assembled system to test again.
What does it do when you specify 11x? Maybe try https://crystalmark.info/en/download/#CrystalCPUID to compare. The BIOS should not matter for this small issue, it is not like the BIOS is intercepting and changing the multiplier request code send from SetMul to the CPU..

5a796d wrote on 2024-10-20, 15:44:

For information, why can't you lower the multiplier below 4x with K7 CPUs?

K7 Should allow 3x, but not 3.5x

for the multiplier issue it is not important, however I will do some tests as soon as I finish reassembling the system.

my mistake, the multiplier reaches 3x

Is there any trick to lower the frequency further?

mockingbird wrote on 2024-10-22, 03:53:

This has to do with the L6 bridge on the top of the chip and not with SETMUL. […]
Show full quote

5a796d wrote on 2024-10-20, 15:44:

the only problem I found doing tests is that if I try to restore the original multiplier (12.5) without rebooting, it sets it to 11

This has to do with the L6 bridge on the top of the chip and not with SETMUL.

See here for info... If you find a safe way to cut closed bridges, please let me know, I am also interested in doing this.

I could be wrong and there may be a way of forcing SETMUL to ignore the L6 bridge. @gerwin - can you weigh in on this?

I avoided the part about cutting the L6 bridges for the same reason as you, I wouldn't want to damage it. the solutions that come to mind for cutting them are with a laser or with a very fine blade

mockingbird wrote on 2024-10-22, 03:53:

This has to do with the L6 bridge on the top of the chip and not with SETMUL.
See here for info... If you find a safe way to cut closed bridges, please let me know, I am also interested in doing this.
I could be wrong and there may be a way of forcing SETMUL to ignore the L6 bridge. @gerwin - can you weigh in on this?

Oh yeah. I forgot about the L6 bridges in my previous reply. Indeed I had to cut one or more bridges. See quote from 2019 below. (And if I had a software trick to do the same, I would not have done so, but I don't have a software trick.)

gerwin wrote on 2019-01-04, 00:29:

...To test it I modified my mobile Athlon processor with a small cut in the L6 bridges: Now multipliers up to 23.0x are accepted for testing.

5a796d wrote on 2024-10-22, 04:25:

my mistake, the multiplier reaches 3x
Is there any trick to lower the frequency further?

3x is quite low in my view, compared to the stock K7 speeds. I don't know about any further tricks for the K7 multiplier. SetMul can only toggle features which were already embedded in the CPU hardware, as-is. Of course it is useful, to set the FSB speed low as well, because it affects the net CPU speed. But still, the K7 platform unfortunately cannot compete with a Slot-1 based system for slowdown options in the critical ranges. Like around 100..150MHz (Deschutes core), or even lower (VIA C3).

gerwin wrote on 2024-10-22, 14:10:

Oh yeah. I forgot about the L6 bridges in my previous reply. Indeed I had to cut one or more bridges. See quote from 2019 below. (And if I had a software trick to do the same, I would not have done so, but I don't have a software trick.)

Thanks for confirming. Any tips on a good method of cutting the L6 bridges?

[/quote]
3x is quite low in my view, compared to the stock K7 speeds. I don't know about any further tricks for the K7 multiplier. SetMul can only toggle features which were already embedded in the CPU hardware, as-is. Of course it is useful, to set the FSB speed low as well, because it affects the net CPU speed. But still, the K7 platform unfortunately cannot compete with a Slot-1 based system for slowdown options in the critical ranges. Like around 100..150MHz (Deschutes core), or even lower (VIA C3).
[/quote]

I understand, I've already lowered the fsb to 100mhz, unfortunately the k7 won't let me lower it less than that, I think it's the fault of the ddr which doesn't work at lower frequencies, but that's already fine, if I could also find some sort of cpukiller for dos or possibly win98 I could lower the frequency even further 😉

mockingbird wrote on 2024-10-22, 03:53:

This has to do with the L6 bridge on the top of the chip and not with SETMUL. […]
Show full quote

5a796d wrote on 2024-10-20, 15:44:

the only problem I found doing tests is that if I try to restore the original multiplier (12.5) without rebooting, it sets it to 11

This has to do with the L6 bridge on the top of the chip and not with SETMUL.

See here for info... If you find a safe way to cut closed bridges, please let me know, I am also interested in doing this.

I could be wrong and there may be a way of forcing SETMUL to ignore the L6 bridge. @gerwin - can you weigh in on this?

however the question of L6 bridges is for those who want to increase the multiplier, which I consider useless because I'm not interested in overclocking

5a796d wrote on 2024-10-22, 18:30:

I understand, I've already lowered the fsb to 100mhz, unfortunately the k7 won't let me lower it less than that, I think it's the fault of the ddr which doesn't work at lower frequencies, but that's already fine, if I could also find some sort of cpukiller for dos or possibly win98 I could lower the frequency even further 😉

IIRC, there does exist a KT133A board with a clockgen that supports 66Mhz... But I have no idea if the VIA southbridge would allow it... Maybe I'll ask the author of VIAFSB to try to implement it.

5a796d wrote on 2024-10-22, 18:33:

however the question of L6 bridges is for those who want to increase the multiplier, which I consider useless because I'm not interested in overclocking

There is a use case for high multipliers --- if you want to run your KT133A Barton at 100mhz FSB. I have done this (20 x 100Mhz FSB = 2Ghz). You put less stress on the northbridge (less heat too) and there is only a nominal performance decrease.

mockingbird wrote on 2024-10-22, 15:35:

gerwin wrote on 2024-10-22, 14:10:

Oh yeah. I forgot about the L6 bridges in my previous reply. Indeed I had to cut one or more bridges. See quote from 2019 below. (And if I had a software trick to do the same, I would not have done so, but I don't have a software trick.)

Thanks for confirming. Any tips on a good method of cutting the L6 bridges?

Pretty sure I used a small Snap-Off Knife. Somehow managed to position the knife just right and push in. Tried to make a picture of the modification, I can barely see it. (attached, Athlon XP-M 2800+ 2133MHz AXMA2800FKT4C).