Apparently these niche CPU upgrades were produced in really large numbers - like lower tens per model.
That's why they are so easily obtainable today.
😀

Found another interesting thread on vogons related to the the same 2x33MHz CPU model:
The Systempro; Benchmarking (TI486SXL2-66)

---

I am able to run SXL2-50 at 60MHz, but not a mm higher.
The system is not fully stable, but Doom hits ~20 fps, when it completes.

Looking at all numbers around - 2x40 seems to be the best overall option.

pshipkov wrote on 2021-04-19, 21:34:

Found another interesting thread on vogons related to the the same 2x33MHz CPU model:
The Systempro; Benchmarking (TI486SXL2-66)

That's one of the 3 I was referring to, but the user kinda butchered it with those bodge wires shown here: The TI486SXL2-66 (PGA-132)... & repair!

pshipkov wrote on 2021-04-19, 21:34:

I am able to run SXL2-50 at 60MHz, but not a mm higher.
The system is not fully stable, but Doom hits ~20 fps, when it completes.

On my chips, I determined that for full stability, 55 MHz is max, but the FPU will suffer at 27.5 MHz. The IIT x2-50 MHz FPUs I tested wouldn't do 55 MHz, unfortunately. Do you recall if the max average frequency that Cyrix FasMath chips can do?

pshipkov wrote on 2021-04-19, 21:34:

Looking at all numbers around - 2x40 seems to be the best overall option.

Yeah. Custom interposer anyone?

I tried several FasMath FPUs - none of them tick reliably beyond 50MHz.
The test i mentioned was performed without FPU.

Stumbled upon another 286 motherboard from Amptron based on the very fast VLSI 20#-16 chipset.
New old stock. Crisp in the antistatic bag.
Soldered crystal oscillator socket. Inserted SIMM slots in the SIPPs for testing.
Works as expected. Gets to 25MHz easily. Didn't check if it can do more.
Basically - similar characteristics as the other ones showcased in previous posts.

Came with 4x1Mb 80ns SIPP RAM modules that don't work reliably beyond 16-20MHz.
There was a 16MHz rated Siemens 286 CPU in the socket. While it worked well at 16MHz, the heat dissipation was extreme. Haven't seen this kind of stuff before.
25MHz Harris or Intersil ones running at 25MHz or higher get warm, but nothing like this.

Here is the guy:

Kind of nothing new to see at this point, except ...

... the box it was in:

The name on the box gives partial answer to the question "Who made these motherboards back in late 1980, early 1990 ?".

Amptron International Inc. was one of the manufacturers.

I'm pretty sure Amptron was a PC Chips branding in the US.

Didn't know that.
I heard the company name, but that's it. Never owned hardware from them. At least not consciously.
There is a chance some of the old junk that passed through here is their doing ...
Searched online - only weak references came up.

If you search for PC Chips M919 you may find references to them being sold as an Amptron DX9700.

Right. Remember seeing references for it around.
Was tweaking today a proper implementation of m912 with those cool pcchips stickers.
Pcchips for everyone !

pshipkov wrote on 2021-04-24, 22:14:

Stumbled upon another 286 motherboard based on the very fast VLSI 20#-16 chipset. New old stock. Crisp in the antistatic bag. So […]
Show full quote

Stumbled upon another 286 motherboard based on the very fast VLSI 20#-16 chipset.
New old stock. Crisp in the antistatic bag.
Soldered crystal oscillator socket. Inserted SIMM slots in the SIPPs for testing.
Works as expected. Gets to 25MHz easily. Didn't check if it can do more.
Basically - similar characteristics as the other ones showcased in previous posts.

Came with 4x1Mb 80ns SIPP RAM modules that don't work reliably beyond 16-20MHz.
There was a 16MHz rated Siemens 286 CPU in the socket. While it worked well at 16MHz, the heat dissipation was extreme. Haven't seen this kind of stuff before.
25MHz Harris or Intersil ones running at 25MHz or higher get warm, but nothing like this.

Here is the guy:

Kind of nothing new to see at this point, except ...

... the box it was in:

The name on the box gives partial answer to the question "Who made these motherboards back in late 1980, early 1990 ?".

Amptron International Inc. was one of the manufacturers.

Very nice. I assume 25MHz was with 1-wait? Or did you find some SIPs that are fast enough to handle 0-wait?

I insert SIMM slots into the SIPP ones. They fit well and contraption is pretty solid.
Then roll with trusted simmsticks = 0 waits.

pshipkov wrote on 2021-04-25, 15:19:

I insert SIMM slots into the SIPP ones. They fit well and contraption is pretty solid.
Then roll with trusted simmsticks = 0 waits.

Nice hack! 😆

It is a nice trick that many use.

---

Last weekend i went through bunch of AMD 486 DX5 CPUs as part of a slowly evolving operation aiming to clean house from hw i dont need.
PC-Chips M912 v1.7 motherboard happened to be on top of the retro hardware pile and was already configured for DX5 processors from previous tests where it behaved itself pretty well, so it was used as a test bed.

First CPU in. Run a mandatory DOOM test ... and the recorded path of the player was so incorrect that seconds after start the action was in a completely different area of the level.
Problems like that are usually attributed to tight bios timings and/or flaky L2 cache and/or flaky RAM.
This was unexpected because never had problems with this motherboard and the usual set of components i use for testing.
Spent good amount of time swapping different sets of trusted sram/ram but no luck for the day.

Next day.
Same board, same components. Plugged one of the DX5 CPUs - everything works. Hmm ...
Can this be a CPU related issue ?
Quickly went through the processors one by one and indeed - an AM486 FX5-133WBGC model (notice the W letter) didn't behave properly.
Tried it on another motherboard - all was good.
This piqued my interest.
Searched online for related info.
apparently m912 is a well known board. Several vogons threads touch on jumper settings and models with fake L2 cache.
No perf data available. No word about CPU specific issues.
Decided to examine it more thoroughly.

It is a new old stock - squeaky clean, no wear.
Despite the ugly Pc-Chips stickers hiding an UMC chipset, i like it a lot.
Late ISA/VLB model.
Supports wide range of CPUs.
72 and 30 pin SIMM slots, maximum of 128mb RAM.
Up to 1024Kb L2 cache.
Great PCB layout. All sorts of info is printed on it, but the DX5 CPU settings are wrong incorrect.

As far as i remember the original BIOS was from year 1994, stored in 256Kb EEPROM from Philips.
Replaced by 1995X code burned on 512Kb chip.

Base metrics in SpeedSys, using an ISA IDE controller.
With VLB EIDE disk i/o is in the range of 10Mb/s.

Perf testing started on a high note - all jumpers and BIOS settings on max for best performance.
Results from interactive DOS graphics benchmarks were not bad with the Ark1000VL video adapter.
Not the best out there, but pretty ok indeed.
Next up - stability/performance in complex computations and Windows.
That didn't go well.
Had to significantly increase wait states for the 3D rendering tests to complete.
Performance dropped to a point where i just lost interest in further testing.
So sharing the DOS graphics results only.

benchmark results

No issues with any of the other AMD DX5 CPU models - ADZ, ADW, VBGC.

Certainly liked the board more before examining it.

Just when i thought that my 386 curiosity was mostly satisfied i had to stumble upon what appears to be Young Micro Systems VEGA 3-486 CACHE hybrid motherboard. It looked very promising, which made me wonder if it can bring a surprise.

The "Made in USA" stamp on it made me smile initially.
My presumption was that the only thing made in USA from this motherboard was the stamp itself, but search of "Young Micro" returned what appears to be a small California based company in the early 90ies, which increases the possibility of the stamp telling the truth.
The board itself was in really good condition, but when i turned it on weird things started happening.
It would lock or go dark unexpectedly. After some time it became clear that this is not an electrical problem, but a physical one.
Applied the trusted bend+twist scientific approach. Things got much better, but after a while i got few more blackouts.
Second session of applied science fixed it for good. Ticking solid now.

Briefly about the implementation:
Symphony Haydn chipset. This is the "old" revision consisting of 3 chips (there is a later variant with 2 chips only).
386 and 486 CPU sockets.
486SX support of up to 50MHz <- this frequency is what sparked my interest
maximum of 32Mb RAM
256Kb L2 cache

With the initial instability out of the way, i moved onto testing its 386 implementation. Didn't bother to check the 486 side.
When on 386 path the board can hit maximum of 45MHz , but is not stable.
RAM, SRAM, VGA and IDE components that i test with can operate at 55MHz no problem.
Limitation is the motherboard itself.
Stepped down to 40.
Also, it does not like SXL2 and DLC CPUs. Tested with 386DX.
Finally, DRAM wait states had to be set to 1 (up from 0).

Here is what SpeedSys says:

At 40Mhz we don't get a performance beast, but things are pretty up there on a clock-to-clock basis.
Compared it to PC-Chips M321, which is another one of those great boards that just don't scale well past the 40ies.
Added and the common by now 50MHz DTK system to indicate peak perf.
Didn't bother with 3D rendering tests. They take forever on a config like this and won't reveal anything shocking.
So, going with lightweight interactive DOS graphics only.

benchmark results

Overall pretty good retro hardware.
It is 50MHz rated for 486SX CPUs, but the 386 path seems to be on a different opinion - too bad.

In series of really unfortunate events two Quadro FX4000 AGP video cards were lost in action - regular and SDI versions.
The cause ?
Faulty EVGA 500W PSUs, no older than 2 years.
Bummer.

Both cards had BIOS version 5.40.02.21.04.
A low-priority item on my to-do list was to try newer version one day.

That day force-happened last week, when standard dual-header replacement QFX4000 ended up in the P2B-D based system from few pages ago.
It came with BIOS version 5.40.02.38.11, which slightly improves performance.
For example - until now Quake 3 timedemo (1570/150,1024x768, all settings on max) hit 258 FPS.
Now - 262 FPS.
Not bad.

This may be of interest to other proud owners of AGP Quadro FX4000 / GeForce 6800Ultra (or GT), so decided to share the info.

Several pages ago (around the beginning of this year) i went through bunch of dual Slot1/2 (Pentium III) motherboards in attempt to identify the best ones and build a badass graphics / software_development workstation.
This initiative is reflected in the next posts:
Supermicro S2DGE, Tyan Tiger-133, Asus P2B-D
OR840 (IBM Intellistation M Pro, Type 6868)
OR820 (Dell Optiplex G300)
(there were 2 more BX440 boards from Gigabyte and Soyo that were discarded early on)

The last one from the OR800 series was an interesting case and took longer, because i put quite a bit of effort to work around some issues with it.

Asus P3C-D (rev. 1.13-A06) / HP Kayak XM600 - a bitter-sweet experience.

Where to start really ?

Assembly is of very high quality.
Standard ATX layout, components and connectors - something that is worth mentioning, as most dual CPU OR8## implementations out there are weird - unusual formats, custom connectors, require specific PSUs, etc.
On-board UltraDMA/66 controller, Crystal CS4280 PCI audio, IR.
AGP Pro slot.
Clock generator that supports up to 180MHz.
That clock gen suggests that the system is meant for overclocking, something that is common for many Asus products and something that i badly wanted to try, but ...

... the HP engineers who built it into the Kayak XM600 were on very different opinion.
All OC-related settings are omitted from HP's customized BIOS.
DIP switches cannot go past 133MHz FSB, regardless of what ASUS's manual says.
Custom HP POST process forces memory check that takes very long time with 1Gb of RAM installed. It really tests your patience.
And so on.

So i spent time trying to circumvent the imposed limitations.
Unsoldered some the custom HP chips and tried to rewire the circuits. That never worked. It was a shot in the dark to start with.
Installed a second DIP switch package, where the default one (from Asus) would have been. That didn't change the resultant behavior of the switches.
Socketed the EEPROM, burned the original ASUS bios on another chip, but board didn't light up with it.
Tried to force higher FSB by hard-wiring the clock-gen. That didn't go well too.
Basically, i was beaten back at every step and 2 weeks ago officially surrendered.
Disappointing.

So, going with the base package that HP thinks is what i deserve:
2 x Tualatin-S 1400/133/512
Quadro FX4000 (NT4) and Quadro2 Pro (DOS)
2x512Gb 40ns RIMMs + terminator, at base 100MHz speed.
3COM LAN card.

In this configuration - no complains - everything works as advertised.

HP's control panel.

Here is what speedsys says:

Looks good overall. Disk i/o is not that great.

The standard set of benchmarks.
Compared to Asus P2B-D at the same CPU/FSB/AGP frequencies and video cards.
Gray bars represent peak perf with P2B-D at 1575/150/48.

Couple of things to notice:
- Global Illumination rendering test, C++ compilation test - P3C-D is clock-to-clock faster than P2B-D.
- OpenGL Quake3 test, scanline rendering test - clock-to-clock things are on par. Keep in mind that the scanline rendering test is heavily influenced by disk i/o (large shadow maps). I used the built-in UltraDMA/66 controller for the tests instead of Promise ATA/100. I am pretty sure that the 5 seconds difference would be erased by ATA/100 in the loop.

The first link in this post touches on turning P2B-D/BX440 into something that is unmatched by any other P3 system.
It takes next-next gen P4 hardware to outdo it.
If HP didn't butcher the P3C-D board - it could tell a story with potentially unexpected end for BX440.
Too bad.

---

OR800 hardware had limited market window back in the day, so Vanilla Asus P3C-D motherboards are rare today. Never saw one for sell really.
Online searches don't reveal anything conclusive about their qualities.
Curious how well they overclock.
Any hints ?

This post is the final chapter of the effort to build a high-end dual slot Pentium 3 workstation from the period 2000-2001.
Choice of components and final assembly reflect my preference towards highly optimized systems for a given mission.

After checking bunch of hardware covered in these previous posts
Supermicro S2DGE, Tyan Tiger-133, Asus P2B-D
OR840 (IBM Intellistation M Pro, Type 6868)
OR820 (Dell Optiplex G300)
OR820 (Asus P3C-D / HP Kayak XM600)
this is what ended in the final assembly:

motherboard: Asus P2B-D, revision 1.6, stepping D03, ICS 9250CF-08 clock generator, HIP6019BCB voltage regulator, VTT mod, VIO mod, 4th jumper mod, ACPI fix, latest BIOS with Tualatin microcodes.
cpu: 2x Tualatin-S 1.4GHz in Asus S370-DL rev. 1.02 slotkets.
ram: 256Mb at 1575/150/37.5 (Micron 2-2-2), 768Mb at 1470/140/35 (Elpida 3-3-3), 1Gb at 1400/133/33 (Elpida 2-2-2)
video: Quadro FX4000 (period correct choices Quadro2 Pro / Fire GL2)
ide: Promise Ultra133 TX2
sound: Sound Blaster Audigy EAX
lan: 3COM 3C925-TX
storage: Sandisk 50Mb/s CF card + CF-IDE adapter
os: Windows NT4 + SP6a (128-bit) + DX5 + IE6

Performance parameters and other details have been covered in previous posts already - wont repeat them.
As a brief reference that will make more sense to the most vogoners - the system running at 1575/150/37.5 achieves the next results in Quake 3 (highest settings):
262 fps at 1024x768
233 fps in 1280x1024
Which is pretty good. At least i haven't seen higher scores posted around.
Anyway, enough bragging.

Here is how things started and ended up looking in the physical world.
Custom 3D printed ATX PC case - exactly to my liking. Spent the time to make it real tight and sturdy.
DIMM holders. I often switch between different FSB speeds to test things. This requires adding/removing sticks to the DIMM slots. Figured i need a place to keep them when not in use. Added one more holder after taking the pictures.
The shown here video card is dead now. Replaced with another Quadro FX4000, non-SDI version, but dual head one.
Photos show the initial cable management which was not bad, but further improved it recently for better optics.

And that's it.

That is an awesome dual Tually build and I love your case it's sweet.

Hey, thanks.

After combing through a lot of 486 motherboards myself and checking what others say online, it looks like the fastest classic (VLB/ISA only) 486 motherboard around is Asus VLI-486SV2GX4 ...
How things changed over time.
Back in the day - one of many.
Today - in a class of its own, for those who care to notice of course.

But nothing is set in stone and there is always a chance for one of the pretenders out there to disrupt the "cosmic order".
Recently stumbled upon very promising looking motherboard from Young Micro Systems, model mb-sw486g-3vl-v11, based on the latest incarnation of the Symphony Haydn SL82C491/2 chipset.
I had great experience with other boards based on this chipset, so decided to spend the time to examine how this one behaves.

PCB is in great condition.
It has 5V only socket. No voltage regulator on board.
(It looks like there was another version with VRM.)
Because of that decided to test with POD100 first, then checked with Am486DX5 CPU sitting on a 5V-3.3V adapter.
AMI BIOS with the clunky windows UI.
All settings on max.
Requires 15ns or lower latency L2 cache for 1-1-1 burst read/write cycles.
Very picky about L2 cache chips but was able to find a set for 1Mb buffer.
Handles well P24T (POD100) L1 cache in in WB mode. Something that is not always the case with other boards.
System is not completely stable past the 40MHz base frequency with 486DX5 cpu. Tried hard but no luck.

Happen to have a second motherboard like this. Somewhat different version, but cannot find any details about it.
Performance and overall behavior is identical for both.

For complete stability at 40MHz FSB 2 BIOS settings had to be set below their optimal values:
BURST READ CYCLES 2:4 = X-2-2-2 (best is X-1-1-1)
BURST READ CYCLE 1 = 3-X-X-X (best is 2-X-X-X)
Also, BURST CACHE FILL must be set to DISABLED for improved performance.

Used Holtek EIDE controller which delivers 10mb/s disk i/o with the VLI board.
Ark1000VL was used for the DOS tests.
S3 Trio64 VLB for Windows GUI tests.

--- AMD Am5x86 at 160MHz

benchmark results

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

Speedsys:

benchmark results

At 16-bit colors the system gets very unstable. Considered that as a failure instead of searching for other VLB cards that may potentially work.

---

Not the best performer, but overall pretty ok.

Coincidentally, not far ago i checked another Symphony motherboard from Young Micro Systems here.
While both boards are actually pretty good they don't take the most out of the otherwise very potent chipset.
The difference is stark when compared to Symphony Haydn systems from DTK (here and here).
So, taking on Asus VLI will wait another day.

Wanted to find the answer to an old question - what is the relative performance difference between Pentium 3 Coppermine and Tualatin CPUs.

So lined-up the next processors with similar characteristics:
SL5BS 900/256/100/1.75V - Pentium III Coppermine
SL5LX 900/128/100/1.75V - Celeron Coppermine
SL5GN 1200/256/133/1.475V - Pentium III Tualatin
SL68V 900/256/100/1.475V - Celeron Tualatin

Overclocked the 900/100 ones to 1200/133, so all CPUs where normalized to the same CPU/FSB frequencies.

SL5BS required 2.1V for complete stability.
SL5LX needed 2V.
SL68V required soldering of adaptor.
SL5GN was a bit more problematic.
It also required soldering of adaptor. During that work i almost ruined the CPU, but everything turned out well at the end.
Then it didn't like any of the Asus S370-DL slotkets for some reason.
I have 7 other ones of various types. In some of them it booted, worked fine in DOS, but hung in Windows.
After a bit of for and back i realized that it really does not like any voltage above 1.475.
Luckily i have this Gigabyte GA-6R7PRO slotket that covers wide range of voltages 1.3-3.5 with a step of 0.05.
Set it to 1.45V - turned out great.

Constant test environment - Asus P2B-D, Quadro FX-4000, Promise TX-100, Windows NT4 SP6. Same BIOS settings, etc.

Here is what i found:

1                        Doom           Quake 1          Quake 3
2Coppermine Pentium 3    122            174.2            137.5
3Coppermine Celeron      119.5          165.8            115
4Tualatin Pentium 3      122            183.9            147
5Tualatin Celeron        122            183.1            147

Doom test cares only about MHz.
Quake 1 test shows around ~5% performance difference between the Coppermine P3/Celeron processors.
Quake 3 test shows ~20% performance difference between the Coppermine P3/Celeron processors.
Tualatins are on par in all tests, which makes sense. They are the same thing really. Both are faster than P3 Coppermine.

Ran couple of heavy offline computation tests in NT4 that won't be of much interest here.
Results were inline with the Quake 3 ratios.

Obviously the Tualatin core is more efficient than Coppermine.
The gap widens further if we consider the server variant Tualatin-S (outside of the scope here).

NOTE: Shared numbers don't show peak performance, but the relative difference between classes/types of processors.