I need glasses.
PVI but not VLI.
Sorry about the confusion.
CoffeeOne cleared it up.

pshipkov wrote on 2023-09-24, 16:29:

I need glasses. PVI but not VLI. Sorry about the confusion. CoffeeOne cleared it up. […]
Show full quote

I need glasses.
PVI but not VLI.
Sorry about the confusion.
CoffeeOne cleared it up.

I thought you got some glasses?

Yes, removing JP18 1-2 is the apparent 5 V setting, however I measure only 4.73 V at the CPU itself (150 MHz), or 4.60 V (200 MHz). 4.98 V is measured at the AT connector. Has anyone tried 200 MHz on the PVI?

How did you overcome the following?

feipoa wrote on 2023-09-24, 08:42:

Some initial thoughts on this board:
I've run some tests on the PVI-486SP3 using a Voodoo3, Am5x86, 256K, and 64M FPM. DOOM benchmarks reveal that this board is faster at 150 MHz (1319 realtics) compared to 160 MHz (1412 realtics), likely because the L2 and RAM throughput's are substantially faster. A major problem I've run into is that I cannot have the Cache Write Cycle set to 2 and use the Voodoo3 card in Windows 95. With this setting at 2, Windows cannot find or cannot assign drivers to the Voodoo3 card. I get a message saying that there is a problem with my display adaptor, which cannot be resolved. If I set the Cache Write Cycle to 3, there are no issues in Windows with the Voodoo3.

Unfortunately, setting the Cache Write Cycle to 3 kills the DRAM performance, e.g. DRAM Read speeds go from 59.7 MB/s to 45.5 MB/s, and DRAM Write speeds go from 104 MB/s to 68.2 MB/s. Whether I am running the system at 160 MHz or 150 MHz, I must set Cache Write Cycle to 3 for the graphics card to work in Windows 95. I don't understand why. I'm not sure if this issue is specific to the Voodoo3, or if other graphics cards are afflicted in the same manner.

What BIOS settings were you using at 160 or 150 MHz in Win95? Did you guys do any testing with a Voodoo3 in Windows?

feipoa wrote on 2023-09-24, 21:13:

I thought you got some glasses? […]
Show full quote

pshipkov wrote on 2023-09-24, 16:29:

I need glasses. PVI but not VLI. Sorry about the confusion. CoffeeOne cleared it up. […]
Show full quote

I need glasses.
PVI but not VLI.
Sorry about the confusion.
CoffeeOne cleared it up.

I thought you got some glasses?

Yes, removing JP18 1-2 is the apparent 5 V setting, however I measure only 4.73 V at the CPU itself (150 MHz), or 4.60 V (200 MHz). 4.98 V is measured at the AT connector. Has anyone tried 200 MHz on the PVI?

How did you overcome the following?

feipoa wrote on 2023-09-24, 08:42:

Some initial thoughts on this board:
I've run some tests on the PVI-486SP3 using a Voodoo3, Am5x86, 256K, and 64M FPM. DOOM benchmarks reveal that this board is faster at 150 MHz (1319 realtics) compared to 160 MHz (1412 realtics), likely because the L2 and RAM throughput's are substantially faster. A major problem I've run into is that I cannot have the Cache Write Cycle set to 2 and use the Voodoo3 card in Windows 95. With this setting at 2, Windows cannot find or cannot assign drivers to the Voodoo3 card. I get a message saying that there is a problem with my display adaptor, which cannot be resolved. If I set the Cache Write Cycle to 3, there are no issues in Windows with the Voodoo3.

Unfortunately, setting the Cache Write Cycle to 3 kills the DRAM performance, e.g. DRAM Read speeds go from 59.7 MB/s to 45.5 MB/s, and DRAM Write speeds go from 104 MB/s to 68.2 MB/s. Whether I am running the system at 160 MHz or 150 MHz, I must set Cache Write Cycle to 3 for the graphics card to work in Windows 95. I don't understand why. I'm not sure if this issue is specific to the Voodoo3, or if other graphics cards are afflicted in the same manner.

What BIOS settings were you using at 160 or 150 MHz in Win95? Did you guys do any testing with a Voodoo3 in Windows?

I cannot say something about Am5x86 @180MHz or @200MHz, I never managed to get such a thing running.

But me experience with the PVI-486SP3 (I have a quite old revision 1.2 but it is already the B-chipset, I use Bios 3.06):
After reading the posts from pshipkov, I tried to use 512kB cache, but it should work with fastest timings @40MHz. At the beginning I thought it is simply impossible.
But finally I found a set of 4 128kx8 srams which work well and a fast tag ram.
So I am able to run with 512kB l2 cache with everything fastest, Win98SE runs, no crashes with the Lightwave 3D program under Win98SE. That and also the Windows 98SE installation is really a good stability test.
The only drawback is the RAM, I can only use 32MB RAM (2 times 16MB FPM). I have 2 sticks of 32MB, but they are not good enough for fastest settings. Also with only one of the 32MB sticks I have to add wait states.

So at 4x40MHz: cache 2-1-2, everything fastest.
Fastest setting is not possible at 50MHz anymore, espcially not with the single bank sram configuration. But as far as I remember I was unable to set 2-1-2 with 256kB double banked sram, too.

I don't have a Voodoo3 PCI unfortunately, I use now mostly a S3 868 PCI with this board. A VLB card would be a bit faster on this board, especially for the DOS benchmarks.

EDIT: I forget to mention the following, but I guess it is already very clear:
I don't try 50MHz anymore, because 50*3 is always slower for me than 4*40MHz. I don't had a single benchmark run, where 50MHz could beat the 40MHz setting.
The same is true for me on the VL/I- 486SV2GX4, also there 40MHz always wins against 50MHz. And as already mentioned 50x4 is completely out of range for me.

Doom was faster for me at 3x50 compared to 4x40.

If you are sure that your cache write cycle was set to 2 and not 3, then I will try some smaller RAM sticks to see if it clears up the issue with the Voodoo3 driver in Windows 95. I'm already using my best 64 MB FPM stick, but from what you've said, it sounds like 64 MB may be too large for tight timings on this board. Nonetheless, I had no issue in DOS.

These glasses are still on their way ...

A while ago, buried somewhere in the depths of this thread, we had a conversation with CoffeeOne about VLI/PVI where we figured out the purpose of the JP16/JP18, respectively.
Until then i was not able to get run Am5x86 processors at 180/200MHz on these boards.
Didn't measure exact voltage to CPU with JP16/18 open, because the systems lit-up properly at 180/200MHz.
PVI was not fully stable. VLI can be stable-ish with added wait states, which defeats the purpose. It is the motherboards themselves.
How i verified - i have a "magical" CPU that runs 200MHz (3x66) at 4V on air. Even with it i was not able to make things work stably.

Doom and a handful of other apps/games prefer 3x150, but most things run faster at 4x40. That's why i also abandoned the 3x150 config.

I run both PVI and VLI boards in 4x40 or 3x50 with all BIOS settings set to their optimal values.
PVI with single bank 512Kb, VLI with 1Mb buffer.
Don't remember trying V3 in the PVI mobo, but have V1 there - seems to be working really well.
VLI is equipped with 64Mb RAM, PVI with 128Mb. Both are FPM.

Thanks for the added information. Could you verify which graphics card you used on the PVI at 4x40, in Win95, using 128 MB FPM, and with the cache write cycle set to 2? I will duplicate your setup to ensure there's nothing wrong with my board.

At 3x50 on PVI, were you also using the tightest DRAM/SRAM timings throughout? If yes, then I'd imagine 3x50 would be faster than 4x40 for most applications that use large amounts of RAM. The DRAM read/write timing differences are just too great.

Is this what your BIOS settings look like? Are you using the SI4I0307 BIOS from 02/11/98?

EDIT:
I've done further testing. With Cache Write Cycle set at 2, I am not only not able to use the Voodoo3 in W95, but even a trouble-free card like an S3 Virge also will not work. Windows says there's a problem with my display adaptor. Setting Cache Write Cycle back to 3, no issue. I tried swapping my 256K double-banked cache for 10 ns pieces, same problem. I tried other (official) 12 ns modules, same issue. I even tried 10ns 512K (reprints), but in this case, I had trouble even getting into Windows. I tried a few sets, same thing. pshipkov's ability to find magic SRAM modules still astounds me.

Next, I tried a VLB graphics card. I used a Trio64, set Cache Write Cycle to 2, and Windows works just fine. Thus, it is highly probable that this issue is restricted to PCI graphics cards. If any of you have a PVI board handy, would you be willing to confirm this observation using a PCI Virge card at 40 MHz?

This isn't necessarily the end of the world, being forced to use a VLB graphics card, since the one VLB slot is kind of the allure of this motherboard anyway. I just wish it had a 60 MHz FSB setting.

feipoa wrote on 2023-09-25, 09:17:

Is this what your BIOS settings look like? Are you using the SI4I0307 BIOS from 02/11/98? PVI_BIOS.JPG […]
Show full quote

Is this what your BIOS settings look like? Are you using the SI4I0307 BIOS from 02/11/98?
PVI_BIOS.JPG

EDIT:
I've done further testing. With Cache Write Cycle set at 2, I am not only not able to use the Voodoo3 in W95, but even a trouble-free card like an S3 Virge also will not work. Windows says there's a problem with my display adaptor. Setting Cache Write Cycle back to 3, no issue. I tried swapping my 256K double-banked cache for 10 ns pieces, same problem. I tried other (official) 12 ns modules, same issue. I even tried 10ns 512K (reprints), but in this case, I had trouble even getting into Windows. I tried a few sets, same thing. pshipkov's ability to find magic SRAM modules still astounds me.

Next, I tried a VLB graphics card. I used a Trio64, set Cache Write Cycle to 2, and Windows works just fine. Thus, it is highly probable that this issue is restricted to PCI graphics cards. If any of you have a PVI board handy, would you be willing to confirm this observation using a PCI Virge card at 40 MHz?

This isn't necessarily the end of the world, being forced to use a VLB graphics card, since the one VLB slot is kind of the allure of this motherboard anyway. I just wish it had a 60 MHz FSB setting.

I did some test runs with my PVI @ 160MHz not so long time ago.
It was 100% stable under Windows 98SE, the graphics card was a S3 868 as already mentioned.
All settings fastest, I think ISA clk was /4 => 10MHz.
Re: 3 (+3 more) retro battle stations

Unfortunately, I do not have a PCI 868 card. But I am also using Win95c on the system, which adds another variable.
Which BIOS are you using and on which MB PCB revision?

I run the PVI board at 4x40MHz, 128Mb RAM, 512Kb L2 cache, BIOS on max, Matrox Millennium 8Mb, Voodoo 1, SB AWE32, LAN, CF card inserted into the on-board IDE header, Win95 OSR2.
3x50 conforms to the same config.
On the same board i tested bunch of PCI/VLB/ISA graphics cards - no problem. Most of them are listed in the first post of the thread.
The magic is entirely in the L2 cache chips and a little bit with the RAM modules too. Some curation is needed. But the board just goes in general.

CoffeeOne was able to achieve tightest BIOS timings for 4x40 as well.
I remember another forum member saying the same in PVI related thread around here.
So it is possible.

Could you let me know which BIOS version you both are you using? What board revision are you using? I am unable to compare apples to apples without further detail. Since this issue is specific to PCI on my board, it is possible there is some degraded component on my board or other incorrect BIOS setting.

Are you using the 3.45V setting or the 3.6V setting for the CPU?

I do not have a Matrox Millennium. I have only the Mystique 220 and Millennium G200. I will try the G200.

I am using CF on the integrated IDE port. No XT-IDE, just an 8 GB CF.

Were you also able to obtain to obtain Cache Write Cycle at 2 using 256K double-banked (40/50 MHz), or is this one of those weird boards which work better with 512K single-banked?

It is curious that you both are not using the VLB port on your PVI board. Isn't this one of the perks of the board?

The Matrox G200 doesn't work on this board. Screen stays blank. Instead, I installed the Matrox Mystique 220 4 MB and its drivers. The symptoms remain:

I tried 33 MHz as well, but the symptoms remain.

I disabled L2 cache entirely, but the symptoms remain.

I tried another 6 types of 50 or 60 ns memory sticks, from 16 MB to 64 MB, but the symptoms remain.

The issue is specific to PCI graphics cards. VLB graphics, no symptoms with cache write cycle at 2. I'm starting to think some issue with the PCI and this BIOS version.

It would be helpful if I know your BIOS version, PCB revision, and chipset revisions. I am using SI4I0307 BIOS from 02/11/98, PCB v1.8, and SiS 496/497 from NV 9617 / NU 9616.

I see that CoffeeOne is using 306 rather than 307. Any reason for the lower revision?

Did you leave the PCI latency timer at 80?

Your explanation brings some memories.
Set CACHE WRITE CYCLE to 3.
You may want to set DRAM WRITE CYCLE POST to 1.
See if that fixes it.

pshipkov wrote on 2023-09-26, 06:22:

Your explanation brings some memories. Set CACHE WRITE CYCLE to 3. You may want to set DRAM WRITE CYCLE POST to 1. See if that f […]
Show full quote

Your explanation brings some memories.
Set CACHE WRITE CYCLE to 3.
You may want to set DRAM WRITE CYCLE POST to 1.
See if that fixes it.

Could you elaborate on these memories?

I've noted a few times here that setting the cache write cycle to 3 fixes things, but it substantially reduces the DRAM read/write performance.

I am trying to figure out how you (and CoffeeOne) are able to run your systems with cache write cycle at 2 on PVI w/PCI graphics?

Just saw your edit above.
Ok, will spend some time tonight in front of the PVI PC and try to remember what was it.

BTW, the 306 BIOS did not help the situation. I see no way around using VLB graphics for maximum performance (Cache Write Cycle at 2). But I'm not sure if other, non-graphics PCI cards will still work, or if having Cache Write Cycle at 2 basically turns the system into a VLB/ISA only motherboard.

The PVI rig is supposed to be my favorite sweet-memories PC.
This implies prefect condition, but its health deteriorated since the last inspection.
Win95 is not fully stable - frequent Kernel panics with apps shutting down.
Will have to spend some time on it over the weekend before answer any questions.
Bummer.

I'm packing my PVI board away now, but before I did, I ran some additional tests. I left graphics on VLB so that I could use CACHE WRITE CYCLE=2. Then I tried to use a PCI network card (Realtek RTL8139D) in Win95c, but the card was undetected. If I set CACHE WRITE CYCLE back to 3, the Realtek network card worked fine. This is unfortunate because it means the issue isn't restricted to PCI graphics card.

I then put CACHE WRITE CYCLE back to 2 and was able to use a PCI Voodoo2 card just fine. So perhaps the issue with PCI and CACHE WRITE CYCLE=2 is related to assigning IRQ resources? To ensure I could still use IRQ's on other non-PCI peripherals, I connected an LPT-to-LAN ethernet adaptor. It worked fine.

So this board does have the ability to match the VLI board in raw L1/L2/DRAM speeds at 40 MHz, but with some noted caveats.

I've spent 4 nights on this board and just could not get CACHE WRITE CYCLE=2 working with PCI graphics or networking. I will be very interested to see how you were able to run your board with CACHE WRITE CYCLE set to 2 all this time.

feipoa wrote on 2023-09-27, 09:18:

I'm packing my PVI board away now, but before I did, I ran some additional tests. I left graphics on VLB so that I could use CAC […]
Show full quote

I'm packing my PVI board away now, but before I did, I ran some additional tests. I left graphics on VLB so that I could use CACHE WRITE CYCLE=2. Then I tried to use a PCI network card (Realtek RTL8139D) in Win95c, but the card was undetected. If I set CACHE WRITE CYCLE back to 3, the Realtek network card worked fine. This is unfortunate because it means the issue isn't restricted to PCI graphics card.

I then put CACHE WRITE CYCLE back to 2 and was able to use a PCI Voodoo2 card just fine. So perhaps the issue with PCI and CACHE WRITE CYCLE=2 is related to assigning IRQ resources? To ensure I could still use IRQ's on other non-PCI peripherals, I connected an LPT-to-LAN ethernet adaptor. It worked fine.

So this board does have the ability to match the VLI board in raw L1/L2/DRAM speeds at 40 MHz, but with some noted caveats.

I've spent 4 nights on this board and just could not get CACHE WRITE CYCLE=2 working with PCI graphics or networking. I will be very interested to see how you were able to run your board with CACHE WRITE CYCLE set to 2 all this time.

I can make more tests.
But CACHE WRITE CYCLE = 2 not even working with a FSB of 33MHz sounds really suspcious.
Maybe you are right and it has something to do with assigning IRQs over PCI. Old graphic cards (also PCI ones) don't need an IRQ assigned. But the Voodoo3 needs it as far as I remember.

I had another strange thing with PVI and Windows 98SE:
When I install it without any PCI card inserted (so using a VLB graphics card and an ISA Nic), there is no PCI bus detected. Changing the graphics card to a PCI one later on creates a lot of troubles. So it is better to install Windows 98SE when a PCI graphics card is installed from the beginning.

Maybe your problem is really only related to Win9x, so not really a hardware issue.

Remembered what was it.
Notice the first 3 parameters in the PCI AND PNP SETUP in BIOS.
video

Ran a basic offline graphics test.
The program often crashes on exit - this is not related to the story here.
It is intended to run on WinNT.
Can work on Win95 as well, but glitches quite a bit.