Hey 386DX40,

had to lookup what PKM-3331 looks like.
Funnily enough the image on the RTW is actually from this post early in the thread.
Apparently the same assembly was sold under different PKM versions.

Good to know you saved the board. These are the fastest 486 ISA assemblies.
In addition to the post i linked above, i remember verifying this statement against SiS 471, 496/497, UMC UM8881#, VIA 495/481. It holds true.
And yes - how this chipset handles ISA frequencies is kind of crazy.
Most Winbond ISA IDE controllers are not great, but you stumbled on one of the good ones - congrats.
The VGA is probably holding you back a little bit. Cirrus Logic GD-5434 ISA cards are very expensive, but ET4000AX (or W32i) are still widely available for reasonable prices, or you can try WD90C33-Z - they are very fast too, yet affordable.

You should definitely try 80MHz oscillator for 4x40, or why not 120MHz one for 3x60. : )

Thanks for sharing.

I actually tried a few ET4000AX cards and a few CL5429 cards, and the DFI WD90C31 is faster and has better image quality on LCD displays (no banding). The two CL5429 cards I have (an STB and a Diamond Speedstar Pro) don't like the 16.5MHz ISA speed and they both fail to POST BIOS reliably. I also have a Paradise branded WD90C31 that is just as good as the DFI card for image quality, but it is maybe 1-2% slower then the DFI card. If you read some old Vogons threads there are discussions that the WD90C30/31 ISA cards are quite overlooked for speed. I'd love to try a CL5434 but I'd spend more on one of those then I have in most of this system 🤣.

My I/O controller is a DTC 2280. After trying several 1994-1996 Seagate and WD 1GB-2GB hard drives, the current WD that was made in May of 1998 was the key to stable 16.5MHz ISA bus operation in both Windows 3.1 and DOS. So part of me wonders if on an ISA system with overclocked bus, the hard drive is more of an issue then the controller card itself?

Yet at the same time the ESS 1868 and 1869 cards I have here both work perfectly fine at any ISA speed I can set this board to. I have no other ISA sound cards to experiment with.

Another thing is I'm running 4 4MB 60ns 9-chip parity 30-pin SIMMs. I've seen discussion that the Symphony Haydn II can do memory interleaving, but I am not clear on whether this occurs with 4 SIMMs, or do I need 8 SIMMs and both banks full? If so I may need to source 4 more SIMMs.....

Here is a picture of my current 'best' ISA video card, and a picture of the system. I'm still doing some repairs on the front panel plastics and the LED speed readout, so I don't have the system fully completed yet.

that is very interesting indeed, about the WDC VGA card.
i have wd90c30 isa and wd90c33-z vlb models and as you said, they are excellent dos blasters, but slightly slower than et4000ax.

i still have this motherboard.
if you confirm that the numbers you posted are the best you were able to achieve with the vga in question, i can match the config but strap an et4k ax card to see what comes out of it on my end.
it will be cool if we have some sort of revelation here. : )

---

didnt observe perf diff with half vs all banks full with this board. dont think mem interleaving is a thing.

---

mechanical HDDs are certainly a factor when overclocking the isa bus.
i remember reading with comprehension why that is, but cannot remember anymore.

I have Diamond SpeedStar 24X (WD90C31A) and a SpeedStar 64 (CL-GD5434) ISA cards and the WD card gives me the best image on LCD displays, so that’s what I use in my 386. I can’t remember if there was a performance difference (but it’s running a stock clock speed anyway).

Right.
I cannot remember what picture the two WDC cards here produce on the 1600x1200 20" Dell LCD i use for retro computing activates, but GD-5434 can make stripes in some DOS games/apps.

pshipkov wrote on 2023-11-24, 21:46:

that is very interesting indeed, about the WDC VGA card. i have wd90c30 isa and wd90c33-z vlb models and as you said, they are e […]
Show full quote

that is very interesting indeed, about the WDC VGA card.
i have wd90c30 isa and wd90c33-z vlb models and as you said, they are excellent dos blasters, but slightly slower than et4000ax.

i still have this motherboard.
if you confirm that the numbers you posted are the best you were able to achieve with the vga in question, i can match the config but strap an et4k ax card to see what comes out of it on my end.
it will be cool if we have some sort of revelation here. : )

The two ET4000AX cards ISA cards I originally tested with many months ago when I first got this system going may not be the best of the best. One was an Everex Viewpoint NI with only 512kb of 80ns memory maximum, and the other was marked "super VGA" and it too only had 512kb of 80ns memory with 4 open slots for upgrade chips. And I no longer have either card. So if you have a high end ET4000AX it would be interesting for you to try your board and see how we can compare performances.

Here is a picture of my AMI BIOS setup screen showing the settings I have found to work best for maximum stability and performance at 16.5MHz ISA. I'm using the AMI BIOS core version 070791 originally and kindly provided to me by Vogons user 'Anonymous Coward' which is now available on the retroweb. Anonymous Coward suggested trying the 'Mr. Bios' Haydn BIOS, but I haven't gotten around to flashing it yet so it's possible that there may be more performance available? There are several settings in the AMI BIOS that can't be set for supposedly maximum performance as they cause boot failure even at 8MHz ISA speeds.

Am also wondering what are the experiences here on Vogons with overclocking the various Kingston/Evergreen etc 5x86 upgrade CPUs that have a 5volt regulator onboard? I have heard in the past that the upgrade CPUs don't like being overclocked. Thusly if I tracked down an oscillator for 40MHz I'd probably have to move the jumper on the Evergreen from clock quadrupling to clock tripling which would result in a 40MHz bus and 120MHz CPU speed so may not be worth it really. I also wonder how ISA overclocking on this board would be affected by a 40MHz or 50MHz bus speed? Oh and the L1 and L2 caches only work in write-through mode.

Also here is what my other Paradise branded WD90C31 ISA card looks exactly like. It has a different RAMDAC onboard and doesn't have quite as nice of an image on LCD that the DFI card does with it's Samsung RAMDAC. That card like I mentioned previously is ever so slightly slower than my DFI branded one. This image was borrowed from https://computer-retro.de/Grafikkarten.html

@386DX40

Here is what i see with the same motherboard, same BIOS version + settings, DOS 6.22, Am5x86 CPU 4x33MHz, ...

... using 2 different ET4000AX ISA cards. A very late one (top) and an early one (bottom, that is going to be butchered soon for a higher cause).

If we are not missing something, the ET4000AX card seems to be doing better in these DOS tests.

Wow, thanks for the results pshipkov! That top card is a very nice ET4000AX, Micro-Labs as I recall? I guess I will have to track down a high end ET4000AX card for this system after all, as your results definitely back up that the ET4000AX is still basically the 'king' for ISA speed! The WD holds it own though!

My next step for our information, will be to break out my XGecu Pro T48 burner and try the Mr. BIOS for the Symphony Haydn and see if it actually works with our boards, and if so is there any more performance to be had? I also see a few more Haydn II roms on the retroweb that were not there several months ago when I got my board running. I tried practically every rom there was on our particular DTK motherboard and none of them worked perfectly, as probably to be expected. Therefore the Mr. BIOS could be interesting and the new roms I have found could be interesting. I feel the various BIOSs for this board are somewhat lacking in the ability to tune everything.

I will report back once I try this.......

Btw, with better RAM and L2 chips you can squeeze 0-WS from this board at 33MHz FSB.

I think, when it comes to ISA video cards it goes like this: CL GD-5434 (unfortunatelly they dont work on each and every motherboard), then ET4000AX, then ET4000/W32# (slightly slower than AX in DOS but much faster in Windows), then WDC 90C33-Z, then the rest.
But in DOS the difference between them is basically within the margin of error, more or less.

I tried several BIOSes for these excellent DTK PKM boards. The version you currently use is the best.
In my humble view Mr.BIOS is more trouble than help. Apart from fast soft reset it rarely offers anything useful, but can be an alternative for some specific cases where decent Award/Ami microcodes are not available.
Figured i should mention it since already spent the time on that stuff in the past and can save you the cycles.

Ok nice to know the current AMIBIOS is the best one, so I won't bother burning more EEPROMS. Interestingly, I have a pretty good selection of 15ns cache SRAM chips here, and I did test them with the XGecu Pro then tried to look for 'golden chips' by testing them on another board and yet never was able to get 0-WS stable on the DTK. I suspect it's my 4MB 9-chip parity 60ns Hitachi SIMMs that are the problem as they are fussy in another board I tested them in that has more advanced BIOS timing controls - they simply don't like the fastest timings. Unfortunately all I have is a stack of these same 4MB Hitachi SIMMs. Have read that 3-chip SIMMs a more likely to cooperate with fastest timings then the 9-chip SIMMs. Problem is I occasionally like to run Windows on this system and 8MB is too limiting, thusly why I'm running 4 4MB SIMMs.

I'm going to start looking for an ET4000AX card with a good quality RAMDAC on it and try one just because! But really I'm already beyond happy with the performance of this system as it sits, and it plays more games then I ever thought would possible to play on this level of ISA only hardware! The other upgrade I want to pick up is a Dreamblaster S2 wavetable add-on for my existing ESS sound card, as while I don't mind the ESFM music, in some Doom mods such as Strain, the ESFM just isn't quite good enough.

I still might look for an oscillator that would allow me 40MHz bus speeds though, mostly just to see if I could get 20MHz ISA speeds to be stable!!!!!!! Even if I have to run the CPU at 120MHz, the ISA video is still my most likely bottleneck with this system. I also will update my benchmark results once I find a suitable ET4000 card!

Do you know what frequency is used for testing the SRAM chips with XGECU ?
So far no two motherboards used the same set and order of 28 or 32 SRAM chips.
I can guess about some of the electromagnetic factors at play, like impedance for example, but the empirically proven above fact suggests that there is no device out there that can be used to distinguish between good/bad chips, since there is seemingly no reliable rule to start with.
Over time i narrowed down 2-3 sets of such chips that drastically cut-down the time needed to find working set for a given motherboard, especially when tight wait states are at play.

At 33MHz just about any 15ns set of chips should be fine. Maybe some minor musical chairs (tm'ed by Feipoa term) activity will be needed to make sure things are tight, but in general this is probably your RAM chips. I would start with 4 modules and use the other 4 to rotate things around - there is a chance from the 8 you can find 4 that can handle 0-WS. But, it sounds like you are already satisfied with the overall performance of the rig, so better investment of time will be to finish the setup with Wavetables and other props for completeness.

I remember running the board here at 50 or 55 MHz FSB with 3x CPU multiplier. It was just fine. Pretty sure it can handle 60MHz as well with some extra effort.

The Very late ET4000AX i posted a picture above is not a result of deliberate search/effort - that was on EBay at the time i was shopping.
Many ET4000AX cards are still affordable. They may not be the most recent models, but their performance is the same really.
Also, with 1Mb RAM only, you don't really need a very capable RAMDAC. At least that's my experience.

I'm not quite sure how the XGecu tests SRAM but in a big lot of 15ns 32kx8 Winbond SRAM chips I bought online, it quickly helped me find the bad ones and the ones that tested bad in the XGecu wouldn't allow my other test bench 486 board to even light up. So I think it's probably a "go" or "no-go" type of test. Your strategy for making a few lots of known to be great SRAM chips makes sense seeing as to all the boards you test and the fact you are looking for ultimate stability. I seem to get my ass kicked inserting and removing SRAM chips, so with my luck I'd probably damage something swapping a bunch of chips around looking for the best ones 🤣.

My RAMDAC concerns are because I use an old Sewell Hammerhead VGA to HDMI convertor and connect my 486 to a Samsung 32" 768p TV, so if the card has noisy output or lines/etc it'll look terrible on the 32" screen. My current DFI card actually looks quite nice on the 32" screen. So RAMDAC quality is something I'll be eyeing when I look at ET4000 cards. One thing I've noticed, is if the RAMDAC is in a DIP type package, it's probably not be gonna be good on an LCD.

Once I get an ET4000 I'll definitely rerun the benchmarks and see if we match! And based on your experiences above with 50/55MHz bus speeds, I am going to get an oscillator for 40MHz operation and try it. I wonder if 20MHz ISA is possible on these boards???? I'm going to attempt to find out!

Ok, so same experience with XGecu.
Until i got one of those tools for removing socketed ICs there was quite a bit of collateral damage.

I tested quite a few RAMDACs a while ago. There is a post with info linked from the first page.
Package does not matter, as far as i can tell. I usually look if the chip is socketed. Saves time to put one yourself, if experimentation with different chips is needed.
But, on a 32" TV imperfect picture will be too visually annoying, so i understand your attention to the pixel output.

I cannot remember what BIOS settings i used at 50+ MHz. Started keeping record of that stuff since then, but the info is missing. Can be verified easily.

One of the benefits I noticed with MR BIOS on my PKM 0033 was that it allowed me to use the Cyrix 5x86.

The world of dual Slot 2 motherboards is very small. Like 20 or so small.
If we throw-in few keywords like "standard ATX form factor", "standard ATX power connector", and "AGP", the list gets as short as 10-15 items.
The number shrinks to just 6 or 7, if we add the requirement of "above 100MHz FSB".
All of them are based on Intel's 440GX or OR 840 chipsets, with the OR840 Slot 2 boards potentially problematic for out of specification activities (notes about that below).

One such assembly is Asus XG-DLS rev:1.03 based on Intel 440GX

Takes both 5/12V and 2.8V Xeon III processors with 2Mb integrated level 2 cache.
Up to 2Gb SDRAM.
AGP 4x.
Intel 10/100mbs network adapter.
Adaptec Ultra2 SCSI controller.
Ultra DMA/33 IDE controller.
Capable clock generator that can go up to 150/83MHz. Added 4th jumper to enable the full set of frequencies.
Swapped resistor from pad R230 to R231 for an ACPI fix.
Plenty of space between the two CPU slots. This is important since these processors run hot and appropriate cooling solution is needed which takes space.

At 100MHz FSB and the tightest BIOS settings the system is completely hustle free and just works.
Great and everything but not very interesting by itself since several other motherboards deliver basically the same.
What i wanted to see if i can get past that point by leveraging the more capable clock generator.
That's where i spent time to add the 4th jumper, which turned out to be a premature "improvement".
If i checked the available above 100MHz frequencies beforehand i would have known that.
Basically the multiplier locked 900MHz Xeons max-out at 112MHz FSB and not a megahertz above, or no lights.
It is not the motherboard but the processors themselves.
Using early Xeon models with unlocked multipliers allow for up to 133MHz FSB and stable system, but performance is way below the 1GHz (112*9) peak.
In this configuration with all BIOS settings on max the system is completely stable !

The 900/100 Xeons maxing out at 112MHz is the reason i mentioned at the beginning of the post that OR840 based motherboards are not a good choice.
They seem to have only 100 and 133 MHz FSB options available.
The 100MHz is baseline and uninteresting in general, while 133MHz is out of reach for processors based on 100MHz FSB.

It is worth mentioning that some cool-down time is needed when changing the FSB using the clockgen jumpers.
In all other boards i can simply reset to get the newly selected frequency, but here i have to power down for something like 30 seconds, otherwise the previous frequency is maintained, despite jumper configuration changed.

Components used for testing:
Two 900MHz (900/100) Pentium Xeon III 2.8V end-of-line processors (locked multiplier).
2Gb 4x512Mb Micron ECC modules. ECC memory is required otherwise the board does not lite up.
Quadro 2 Pro, Quadro FX 4000, and FireGL 2 video cards. The FireGL 2 card was unstable in Quake3 at 112MHz. Not sure why. Everything else checked out.
Promise Ultra ATA/133 IDE controller with CF adapter + card. Disabled the SCSI interface. The onboard IDE adapter is nothing to brag about - gravitates around 10Mb/s.
Windows NT4 SP6.

What SpeedSys says. Not bad at all.

And the standard set of tests.
Was not sure what to compare to.
After some for and back decided to go with Supermicro S2DGE (900/100) for the baseline and Asus P2B-D (1575/150, 1470/140) for the peak performance.
Quadro 2 Pro is faster in DOS, Quadro FX 4000 is better in WinNT. Used them that way.

Results are obviously much better than a standard 900/100 system, but slower than 1.4/1.5GHz overclocked dual P3-S'es.
Staring some more at results from previously tested dual Slot 1 motherboards in this thread, i am noticing that Xeon 1000/112 is actually faster than P3-S 1000/133, often by quite a bit.
Looks like Xeon's 2Mb level 2 cache offsets the lower FSB and more. This is somewhat expected because all apps/games/tests from that time are hard CPU bound. Especially considering the used video cards for the tests.

I have to say that i really like this motherboard. It is solid and fast.
The overclocking experience was painless for once. Things just work at up to 1000/112 then hard no go. Easy.
At the same time i cannot but notice how the CPU retention brackets for this motherboard are lower quality plastic and simpler design than what Supermicro offers.
Not a problem really, but leaves an impression.

That is such a beautiful setup! Loving dual slot-2 boards even more than dual slot-1 boards 😁

These are rare items which take effort+patience to obtain.
The few i touched so far are very server-like. Concrete implementations with little room for experimentation.
Still, very interesting in their own way.

My last hope for climbing further than Asus P2B-D is the OR820 based P3C-D with its 180MHz capable clockgen.
Unfortunately the board i have is real quiet ...
Keep butchering HP Kayak M600 derived abominations while trying to convert them to vanilla P3C-Ds. Almost started feeling guilt.
With that said - i think i narrowed down a path to there, but will need one more board (hopefully the last one) to go under the soldering iron.

Couple of notes related to VLB graphics adapters, for the record, and mostly to myself.

1. In continuation to the 486 + 80MHz FSB story.
Ark1000VL VLB card ticking really well at 80MHz in Chicony CH-471B rev.2 motherboard
Keep in mind that the board has a built-in VLB EIDE controller.
So we have 2 VLB devices working side by side at 80MHz just fine.
Nice to see this in action.
Decided to post it here since sometimes we see strong (and incorrect) statements about VLB limitations related to number of devices and higher bus frequencies.

2. Spending top $ on VLB cards with S3 Vision chipsets (964, 968) with less than 4Mb RAM is pointless from usability standpoint.
They are same or slower than S3 Trio64 VLB in DOS and same in Windows.
Can handle higher Win GUI resolutions, but that is only for the 4Mb RAM model, otherwise same as Trio64.

3. There are very few VLB video cards based on WDC's WD90C33-Z chipset. Most of them come with 80MHz RAMDAC. Some rare models (from STB) come with faster RAMDACs. Still, it is pointless to spend the extra $ on them since both the Win 3.11 and Win9x drivers do not offer screen modes that can take advantage of that.

4. VLB video cards based on ET4000/W32 chipset revisions "i" and "p".
Such adapters come with up to 2Mb RAM. The best supported image resolution is 1024/768/64k.
In Win95 both W32i and W32p support it. Apparently the corresponding drivers adjust the MCLK accordingly.
However there are two distinguished drivers for Win3.1. One for W32i and another for W32p. They cannot be mixed.
The W32p driver adjusts the MCLK value to enable 1024/768/64k, but the W32i one does not.
Also, as long as the RAMDAC is 80MHz or faster rated - the above image resolution is reachable just fine = no point spending extra $ on cards with faster than 80MHz RAMDAC.

While there’s talk of ISA cards with WDC chips….I came across this card in a box today. I’ve never seen such a large ISA VGA card before. I’m thinking about using it in my 486 or maybe it would be better in my 386?

zuldan wrote on 2024-01-04, 10:27:

While there’s talk of ISA cards with WDC chips….I came across this card in a box today. I’ve never seen such a large ISA VGA card before. I’m thinking about using it in my 486 or maybe it would be better in my 386?

2F28952B-68FD-4D1A-9A19-8E7C4423503D.jpeg

Amazing card, really old. chips are from 1991.
16 times 64kx4 RAMs with 100ns, so in total 512kB.
No, please don't put it into a 486, it fits more to a 286 or older machine.