Updated my post with the ISA bus frequencies.

I use 50ns rated RAM chips (4x4Mb modules) that work well under much more extreme conditions.
The problem is not the memory itself, but the motherboard.

Simple DOS tests/apps/games work fine at 2 wait states, but the system is not fully stable and the more complex stuff is no-go in some of the runs.
The posted performance metrics reflect fully stable configuration.

Btw, what local storage controller you used when tested the board ?

Adaptec AHA-1522B with SCSI2SD v6.

Didn't expect i will be making this post, but things turned better than anticipated, so here i am.

Caching Technology Corporation 386/486 revision E based on Efar EF8290WB - 82EC495, 82EC392 (which is supposed to be relabeled OPTi 82C392, 82C495B ?)

This motherboard came dead on arrival.
Big battery leak long time ago destroyed a lot of real estate in the upper left corner.
Weak repair attempt at the time was unsuccessful, so i left the board marinating in the graveyard pile.
Until recently.
The inspection of Matra 486SLC2 VESA revealed interesting characteristics of Efar EF8290WB.
At the same time that was an on-rails experience with the on-board soldered CPU/FPU.
That sparked my curiosity about how the chipset will handle wither range of CPUs/FPUs.
Spent the time for proper repair of the Caching Tech Corp motherboard.
Most of the traces were eaten up by the acid to a point that it was not possible to figure out what the connectivity was.
Managed to find few photos online that helped with that.
The fixing of the surface rails led to an interesting visual phenomena - somewhere in the inner layers a short circuit resulted in a "lighting within" - the translucent PCB was glowing in that area from inside.
Was not sure what to do about it.
Felt that i am loosing any chances of fixing the board.
Decided to dig to that layer without further plan.
Luckily the fact of digging and reaching that place was enough by itself to separate the touching rails.
Board came to life. For my great pleasure.
After that spent time to cleanup and apply multiple layers of conformal coating.
End result looks pretty decent.

With the board seemingly working fine i moved onto testing it.

One thing that bugged me from the beginning was the 50MHz crystal oscillator.
Was it a sign of slow-system, assuming that's how the board left the production facility, or it was inserted there later based on some unknown circumstance.
(turned-out to be the first case)

Motherboard works at up to 45MHz FSB (90MHz oscillator). No lights past that.
Unstable past 35MHz (70MHz oscillator) if FPU is present. This was a real bummer.
Decided to perform partial testing at 40/45MHz without FPU instead of going down to 35MHz, which is too slow base frequency and not very interesting to start with.

For the tests i used STB Nitro ISA graphics card with 2Mb RAM (Cirrus Logic GD-5434), 256Kb level 2 cache, 16Mb 50ns RAM (4x4).

--- IBM BL3 at 90MHz (2x45), ISA bus at 11.25MHz

System is inherently unstable regardless of BIOS parameters.
Most tests never finish.
Frequent hangs in DOS.
Impossible to get into Windows.

--- IBM BL3 at 80MHz (2x40), ISA bus at 10MHz

System is fully stable (without FPU).
All BIOS settings on max, except:
DRAM WRITE WAIT STATE = 1 (best is 0)
DRAM WRITE READ STATE = 1 (best is 0)
AT BUS CLOCK SELECTION = CLKI/6 (best is /3)

Clock-to-clock performance is not bad, but at "mere" 40MHz FSB it is at the bottom of the charts.

--- TI 486SXL2 at 45MHz, ISA bus at 7.5MHz

System is fully stable (without FPU).
All BIOS settings on max, except:
DRAM WRITE WAIT STATE = 1 (best is 0)
AT BUS CLOCK SELECTION = CLKI/6 (best is /3)

Speedsys produces corrupted graphics with this processor, which is often the case on many motherboards.

Performance is not great. One of the slowest boards around.

--- AMD 386DX at 45MHz, ISA bus at 7.5MHz

System is fully stable (without FPU).
All BIOS settings on max, except:
DRAM WRITE WAIT STATE = 1 (best is 0)
AT BUS CLOCK SELECTION = CLKI/6 (best is /3)

Again, very low to intermediate at best performance.

---

benchmark results
Look for CTC 386/486.

Kind of disappointing experience overall.
The system is clearly designed for lower frequencies and didn't cope well with the expected 40-45MHz.
Matra's board is a much better integrated assembly that makes the EFAR (OPTi) silicon look much better.

Update for Asus VL/I-486Sv2GX4 @160MHz

Played around with Vesa Local Bus Elsa Winner 1000AVI 2MB. It has the S3 868 chipset.
If someone needs a picture of the card see here 😁
Elsa Winner 1000AVI Vesa Local Bus - Windows 98 problem

So unfortunately Transparent not possible, only Synchronize.
But still the S3 chipsets enables some tools to play with: s3vbe20 (and others) and mclk
For now, I used only

1s3vbe20 linear

and

1mclk /f 2 6 /0 92 7 1

# increases RAM clock from 66.6 to 74.5MHz

1mclk /f 2 6 /1 3

# fast page mode instead of 2 cycle EDO. Hard to believe, but faster setting.
I believe 74MHz is borderline (the card has 70ns EDO), but there are no artifacts.

Dos values:

13D Bench 1.0c                      93.4
2Chris 640x480                      35.2 / 21.1fps
3PCPlayer 640x480                   11.4
4Doom Max                           1233 / 60fps
5Quake                              18.0
6Speedsys 4.78 Video                15129
7Wolf3D                             125.1
8
9Autodesk 3D Studio R3 chevy        1:50 / 110 sec

Windows 98SE Values 800x600 16bit:

1Wintune2                           10826 kPixel/sec
2Lightwave3D Blade                  24:14 / 1454sec

The DOS values are still below Ark1000VL, only PCP 640x480 is higher, because it uses the linear framebuffer.
With transparent this value would be > 12, but that I would consider cheating, because loading of Windows 98SE does not work with this setting.
Also impressive is the 18.0 quake value.
The wintune2 value is also quite good, please note that this was taken with Win98SE.
The lightwave bench is my main stability test, because of the long duration.

Nice.

Some of the results seems to be contradicting each other. For example:
18 fps in Quake 1 suggests well optimized system, but the 3DS test suggests otherwise.
Both of them lean heavily on the FPU.

Wonder if a different VGA BIOS will help the DOS pixel pushing (Wolf3D, Doom) ?
You probably saw this post by @WJG6260, also this one.
Plenty of S3 cards there with details about BIOSes and other stuff.

Do you have the rest of the VLI BIOS settings maxed out ?

pshipkov wrote on 2023-05-29, 02:25:

Nice. […]
Show full quote

Nice.

Some of the results seems to be contradicting each other. For example:
18 fps in Quake 1 suggests well optimized system, but the 3DS test suggests otherwise.
Both of them lean heavily on the FPU.

Wonder if a different VGA BIOS will help the DOS pixel pushing (Wolf3D, Doom) ?
You probably saw this post by @WJG6260, also this one.
Plenty of S3 cards there with details about BIOSes and other stuff.

Do you have the rest of the VLI BIOS settings maxed out ?

I think the system is maxed out. I don't think different bios roms for the S3 would improve, because I overclocked the ram of the card to the max I believe.
3DS value is also top.
I guess you mean LW3D (not 3DS!), that is pretty slow in comparison to your values.
I use Windows 98SE and you Win 3.11, that is definitely a big difference.
But I would really like to know on what does it depend ....

I posted the value 1454 seconds. Today I repeated the test with the same bios settings (the only difference is that I did not execute "s3vbe20 linear" before starting Windows 98SE).
And: I have two times the exact same value: 1464 seconds. So 10 seconds slower.
I searched back in this thread and found this:
Re: 3 (+3 more) retro battle stations
It was 1390 seconds with the Ark. That is more than 1 minute faster than my values from today. WTF. The run with the ark was made with local bus setting "Transparent", can this really make a difference in LW3D? Unlikely.

By the way I don't see that flucutations with 3DS, more less always the same (with same cpu and same mainboard with identical bios settings). But the totall duration is far less and it is running in dos.

EDIT: It must have something to do with the Windows 98 installation. Maybe I try Win 98 lite,

Here is what i meant.
In your last post you reported Quake 1 reaching 18 fps which indicates great overall system performance.
But the 3DS and LW3D results are 150 and 1454 seconds. They should be around 110 and 1290.
This is what i found contradicting.

pshipkov wrote on 2023-05-30, 16:37:

Here is what i meant. In your last post you reported Quake 1 reaching 18 fps which indicates great overall system performance. B […]
Show full quote

Here is what i meant.
In your last post you reported Quake 1 reaching 18 fps which indicates great overall system performance.
But the 3DS and LW3D results are 150 and 1454 seconds. They should be around 110 and 1290.
This is what i found contradicting.

3DS IS 110 seconds.

Doh. You are right. Sorry. So all good there.
LW3D can easily be influenced by win98 vs 3.11/95 that i usually test 486 hardware.

Until now this post outlined the highest performance watermark for 386 class hardware, but nothing lasts forever.

MicroStar MS-3124 version 2.1 based on Contaq 82C591 A, 82C592

The fastest 386 motherboard
... to date.

386 motherboards with this chipset are very rare.
Apart from two dysfunctional boards that i obtained in the past - the one showcased here and another one (no FPU socket), nothing else came up in the usual e-market places over the last several years.
But the recent success of resurrecting Caching Technology Corporation 386/486 revision E (few posts above) prompted me to give MS-3124 a proper try.

There was a lot of damage from a leaked battery.
Most traces in the area were destroyed completely, on both sides.
Close inspection revealed that inner layers are affected as well.
Chances were low, but inspiration was still high.
Was able to piece together the surface rails from couple of photos online.
A bit of guess work for rails under the power connector.
Work took multiple iterations and a serious effort.
This culminated in the system coming to life.
Except that the keyboard was not working.
After multiple rounds with the multimeter and comparison with other boards it was determined that 2-3 traces between the keyboard connector and controller were damaged somewhere in the inner layers.
Bodge wires on the back patched the issue.

I fixed many dead motherboards throughout the years and never thought much about it.
But fixing this and the mobo from few pages above (both rare earths) in a row made me feel like hundred dollars.

With things back online i moved to testing it.
Curiosity was high.

EDIT (several months later):
Sheer dumb luck guided me to another MS-3124 board - this one in perfect health. Both motherboards perform similarly.

It is worth mentioning the easy jumpers setup.
The MSI designers simplified it to 2 resistor arrays and one jumper.
Big +1.

AMISETUP reveals hidden BIOS parameters for ADAPTOR ROM SHADOW, DRAM HIDDEN REFRESH, DRAM SLOW REFRESH, CPU CACHE MODE (WB/WT).
The DRAM ones work at up to 50MHz base frequency and make the system unstable at 55MHz. The CPU CACHE MODE makes no difference with IBM BL3 chips - i think their level 1 cache is operating in write-through mode regardless of what the motherboard BIOS says.

Ok, back to testing.
Used STB Nitro 2Mb video card based on Cirrus Logic GD-5434, standard ISA IDE controller, 256Kb level 2 cache, 16Mb 50ns RAM.

--- AMD 386DX at 50MHz, ISA bus at 12.5MHz

All BIOS settings on max, except:
DRAM READ WAIT STATE = 1 (best is 0)
ISA bus divider = 8 (best is 4)

System is fully stable.
No problems at all entirely.
Feels great.

Intermediate performer.

--- TI 486SXL2

System hangs during POST if level 2 cache is enabled.
Tried all sorts of settings and tricks, but there is some hard incompatibility.
This was unexpected and disappointing.
Tried the 486 BIOS from the other Contaq board i have, but microcodes are incompatible = no lights.

Decided not to test the no L2 cache configuration.

--- IBM BL3 at 110MHz (2x55), ISA bus at 27.5MHz

All BIOS settings on max, except:
SRAM WRITE WAIT STATES = 1 (best is 0)
DRAM READ WAIT STATE = 2 (best is 0)
ISA bus divider = 4

System is completely stable. Very satisfying.
Nothing more to say really.

EDIT:
Well, the above statement was a bit premature.
After series of grueling tests it was determined that instabilities can appear under certain conditions in Windows.
Complete stability was achieved with even more extreme curation of components - the best of the best were needed.
- good PSU with solid 5V line
- magical ULSI DX/DLC 40MHz rated FPU (in addition to the already magical IBM BL3 CPU in use)
- selected 110MHz crystal oscillator (from the tested 5-6 only 2 can do it, the rest made the system flaky, or no lights at all)
- selected keyboard controller chip - KB-BIOS-VER-F by AMI - these are the best for overclocking of 286/386/486 hardware
- this one is extra weird - RAM check during POST should not be interrupted, or the system gets flaky. 100% reproducible, but don't ask how that was determined.
The end result feels great, but the path to there took multiple unexpected turns, described in following posts.

It is worth mentioning that SRAM WRITE WAIT STATES can be set to 0 for tiny bit of extra speed without observable side effects for games and casual tasks, but may affect complete stability under heavy load.
This configuration is reflected in the combined charts (linked at the bottom of this post) with gray bars as an indication for peak performance, but at potentially compromised stability.

And this is how we arrive at the fastest 386 motherboard to date.
Simple as that.

---

benchmark results

Not the fastest clock-to-clock assembly, but overclocks great and is solid.
Once i set the one-two BIOS parameters listed above - the thing just works.
Had to find really good IDE controller, because many of them flake hard with the 27.5MHz ISA bus frequency.

For reference:
My peak performance/stability testing starts with all BIOS settings on max and then relax wait states until pass POST, BOOT and full set of benchmarks.
This is often a convoluted process because i start with quick and easy to pass tests and progress with more and more complex ones.
If the system starts misbehaving somewhere along the way i have to go back and retest, etc.
Not this time.
Basically how DTK PKM-4036 turned-out several years ago.

It looks like another upgrade is coming to the "fastest 386 PC in the west" that started on page 1 and went through several transmogrifications already.

Realized that not all crystal oscillators work on all 386 motherboards.
This is especially true for those 100/110/120 MHz oscillators.
So, i had to go back and recheck all boards that could overclock to 50 or more MHz FSB. Luckily they are not that many and only 2 of them had to be retested.
As a result the above post for MSI MS-3124 motherboard was updated and now states "the fastest 386 motherboard" (to date).
(this may require CTRL+F5 content refresh in case some images are stale in your local web browser cache, which can lead to confusion - discrepancy between text and pics)

---

Updated the post for Asus ISA-386SIQ (SiS 85C461)
User @JakeThompson1 shared some utils for adjusting wait states, ISA bus and few other parameters of the assembly, because the BIOS does not offer any control over them.
This prompted me to do a search for alternative BIOS microcode.
Found one from Soyo SY-019 which worked pretty well and offered enough control.
Long story short: with tight wait-states performance increases by 30% on average which moves the board from very slow (at the bottom of the combined chart) to intermediate.
Not a mindblowing outcome, but at least it confirms that SIS460461 based boards are lacking.
Something that was not clear in the past because of the crippled BIOS.

pshipkov wrote on 2022-12-28, 02:07:

Chaintech 486SPM M104 (seems to be the latest revision), based on SiS 85C496, 85C497. […]
Show full quote

Chaintech 486SPM M104 (seems to be the latest revision), based on SiS 85C496, 85C497.

In crisp condition. Really well preserved.

Feels like slightly early design compared to the very late 486 motherboards.
No support for EDO RAM.
No 3.45V to CPU option.
Very easy to setup jumpers. Very nice.
There is not much more to add - very straightforward assembly.

--- Am5x86 @160MHz (4x40)

All BIOS settings on max.
PnP BIOS auto-config can hang the system during POST. Requires BIOS reset to resolve it.
Turning that parameter off resolves the problem entirely.

SpeedSys:

benchmark results

Not the fastest motherboard, but not bad either.

--- Am5x86 @180MHz (3x60) and @200MHz (3x66)

There are no clock generator jumpers for 60/66MHz. May get back to this with a simple mod at some point later.

--- Am5x86 @200MHz (4x50)

POST completes but always hangs during BOOT to DOS.
It is related to the level 1 cache. If turned-off or chips not present - all is good.
Tried hard to overcome the issue but without success.

--- Intel Pentium Overdrive P24T (POD100) @100MHz

All BIOS settings on max.
No jumper changes required between Am5x86 and POD100 CPUs, which is pretty nice.
Either the motherboard supplies 5V to CPU by itself, or POD100 works fine at 3.3V on this board.

Updated the POD100 comprehensive summary post with the related information for this motherboard.
The short summary:
Intermediate performance in interactive DOS graphics.
Number 3 in Quake 1, number 2 in Windows GUI and one of the top 3 in complex offline computing.
Quite satisfying.

---

The SPM model is a clear upgrade over the previous SOM model which does not handle POD processors and shows overall lower performance metrics.
Also, very easy to work with system.
Too bad it does not offer 60/66 clock generator jumpers out of the box.

The level 1 cache chips cannot be removed from the Am5x86. I think you meant to write "level 2 cache" above, not "level 1 cache". Assuming you meant Level 2 cache, which sizes of L2 cache did you try at 50 Mhz? You mentioned that you had to disable the L2 cache to get it to POST at 4x50 MHz, but did you only try 1024K?

I have my 486SPM modded to run at 60/66/80 MHz FSB. It was a simple mod. I don't have any other data yet, except that I also have funny issues with IDE when PCI = 1/2*FSB.

I got caught up for some time because my VRM was outputting 5 V instead of 3.45 V. The VRM itself went bad and has since been replaced. The VRM had been working just fine when I stored the board in 2015. I bet it would be a full-time job for most of us just to keep our old stock of hardware in working order.

pshipkov wrote on 2023-06-03, 18:40:

And this is how we arrive at the fastest 386 motherboard to date.
Simple as that.

I dunno man, it contains a socket that can accommodate a 486 CPU. Doesn't this make it a 486 motherboard which can also run 386 CPUs? Or maybe a better description would be 'fastest motherboard to run an IBM BL3'.

I wonder if the PGA168 SXL2-66 would work on this board [without breaking it]?

What size and type of wire were you using to repair the traces? I remember when I fixed my Chaintech SiS Rabbit board with about 2-dozen patch wires. That felt nice.

Did you use your scope to ensure that the ISA bus is at 27.5 MHz? That sounds awfully darn fast. Did you test sound, floppy, ethernet, Windows 3.x/95? From my experience, the fastest theoretical ISA bus speed doesn't always translate to the fastest benchmark results.

While I have a brief moment, I wanted to add something about the MS-3124. I’ve had this board for a while and I can indeed confirm the difficulties with SXL CPUs. DLC parts also do not work well, if at all.

However, this board is very strange insofar as it isn’t truly capable of taking a PGA168 CPU. I think the socket it has is actually PGA175. This is a 386/486 hybrid board, but the onboard CPU socket is designed such that it can accommodate the PGA144 Chips & Technologies J38605DX. I believe-but may be mistaken-that this is the only known board that can accommodate this odd CPU. Its main notable feature is the 512 bytes of instruction cache, which the PGA132 J38600DX otherwise lacks. I think that makes it kind of neat, but I’ve no idea what performance looks like. Supposedly the J38600DX is a little faster than Intel/AMD’s parts, and it was a clean-room design, but the 486DLC/SXL are in another league; of course, so is the IBM BL3.

The 486 side of things comes via a daughter card, referenced in the manual as the MS-4901. I don’t think I’ve ever seen one. Does anyone know what these daughter cards look like or how they connect?

I’ve attached a copy of the manual from TRW. The CPU information is derived from Section 2-2, or page 11 of the pdf.

WJG: Nice discovery. I've never seen a motherboard with a socket specifically for the J38605DX - I only have the J38600DX, which fits PGA-132. Since the board lacks any type of official looking CPU riser slot, I guess I will retract my previous statement that its a 486 motherboard. Did you try the BL3 on your board? The SXL issue can probably be resolved upon going through the SXL hardware troubleshooter guide I posted somewhere.

@Feipoa
Fixed the original post. I meant level 2 cache, but not level 1. Good catch.

As for modifying the Chaintech 486SPM motherboard - during testing i was warm on the idea, but list interest soon after that.
That assembly didn't impress, so motivation was not there. Maybe at some point later.

---

That MSI board is a weird beast for sure. It certainly is a hybrid system based on hybrid silicon. That does not bother me much. I am going for BL3 CPU to start with. Gone are the days where i was going for "pure" 386 class machine.

I also noticed the unusual CPU socket, but didn't bring it up, because was not able to find any relevant information, so it was not clear why it is like that.
@WJG6260 no idea where you found the manual, but sharing it here is well appreciated.
Appreciated are and the notes about C&T's CPU and MS-4901 adapter.

I was really impressed by how scalable the motherboard is with BL3 CPUs.

27.5MHz ISA bus - i actually filmed a video of the board going through bunch of DOS/Windows tests with the oscilloscope attached to pin 20 and showing 27.5MHz on screen.
It is kind of crazy, but very satisfying. Decided not to post it at the end - some of the tests take too long which is boring to watch. But can do, if you want to see it.
SB Pro and SB16 work just fine.
Had trouble with PnP AWE32 - it does not recognize DMA. But the same was present when running the system at 33MHz FSB for example. I didn't spend much time on the AWE32, since don't intend to use it for that system. Felt like the usual conflicts at play that one has to spend time on.
I don't test 386 hardware with Windows95/NT, only Windows 3.1x.
Win 3.1 stability ? Solid.
Didn't try yet LAN cards. Will get to that soon - during the transition of the hotrod 386 PC from DTK PEM-4036Y to this MS-3124. I feel it will be ok, that's why not rushing it.
But first i need to print a new case, because the acrylic one i use currently is too small for the MSI mobo.

WJG6260 wrote on 2023-06-19, 14:37:

However, this board is very strange insofar as it isn’t truly capable of taking a PGA168 CPU. I think the socket it has is actually PGA175. This is a 386/486 hybrid board, but the onboard CPU socket is designed such that it can accommodate the PGA144 Chips & Technologies J38605DX.

I think there is another. Although currently the model name escapes me, you may recall there were some full length 386 boards on eBay a few years back based on PEAK/DM that supposedly came out of routers. They also had the PGA144 socket. I think the item listing mentioned “Electroglas”

@pshipkov, did you try 256K double-banked on the Chaintech 486SPM at 50+ MHz? Or did you only try 1024K?

Attached here are some photos showing how to do the clock generator mod on the Chaintech 486SPM. I've also attached the datasheet for the clock generator on this board, an IMI SC464. S0, S1, S2 set the clock rate. Each of these pins have an internal pull-up to Vcc. S0 and S1 are connected via Jumpers JP22 and JP23, while S2 is not connected at all. All we need is a jumper option to set S2 to GND or floating.

I found a convenient GND location here. Notice the solder removed from this GND via:

I will connect one lead of a header to the GND via, the other will go to S2 on the PLL:

With the jumper removed, you get the usual:
FSB, JP22, JP23
25 MHz, 2-3, 2-3
33 MHz, 1-2, 1-2
40 MHz, 1-2, 2-3
50 MHz, 2-3, 1-2

With this new jumper connected, these same positions turn into:
FSB, JP22, JP23
30 MHz, 2-3, 2-3
66 MHz, 1-2, 1-2
80 MHz, 1-2, 2-3
60 MHz, 2-3, 1-2

256K L2 cache yields better timings compared to 1024K. Best I could get with my uncurrated SRAM at 180 MHz was 2-1-3, 3, 0, 1, 1, 2, Slower.

Cachechk
L1 = 186 MB/s
L2 = 91.4 MB/s
RAM read = 52.7 MB/s
RAM write = 83.3 MB/s

DOS Quake = 18.6 fps

However, I did not fiddle about with memory sticks, nor SRAM chips. I did have better luck at 60 MHz compared to 50 MHz though. Odd.

Didn't bother with 256Kb.
When i was examining these boards the objective was to max-them-out in all ways and see where the boundary is.
So it had to be 1Mb level 2 cache buffer.
I am sure 256Kb will be more stable, but that's not very interesting (to me at least).

Thanks for the photos of the mod.
Linked from the original post about the Chaintech 486SPM motherboard.