That's why I inserted (what?!) in my message. It was supposed to be a joke. Heh.

Looks like Chadti99 has the most magical silicon.
3 chips here can handle 180mhz without peltier. Two of them at default voltage.
One can complete all standard benchmarks used around.
Can do the DOS stuff and casual win95 activities.
Cannot survive VC++ compile and other complex compute in Windows - kernel panic or invalid memory errors.
That's the best I got.
lmost practical.

pshipkov wrote on 2022-01-22, 19:10:

That's why I inserted (what?!) in my message. It was supposed to be a joke. Heh. […]
Show full quote

That's why I inserted (what?!) in my message. It was supposed to be a joke. Heh.

Looks like Chadti99 has the most magical silicon.
3 chips here can handle 180mhz without peltier. 2 of them at sock voltage.
One can complete all standard benchmarks used around.
Can do the DOS stuff and casual win95 activities.
Cannot survive VC++ compile and other complex compute in Windows - kernel panic or invalid memory errors.
That's the best I got.
lmost practical.

I’ve only really ran games on the Windows side with this CPU. Let me know something else I can try to be sure.

One of the stability tests i run: https://www.petershipkov.com/temp/retro_pc_im … ev_d_2_long.mp4
Running in parallel:
C++ project compilation (rebuild)
Internet explorer loading some pages - this app is actually really sensitive to instabilities on its own.
3D Rendering
Gl Quake
optionally - start a bulk file copy (this test didn't have much impact on stability, so not running it all the time)

So far nothing survives that without Peltier cooling.

If you want i can zip the apps/projects and pass them to you. Or image of the partition.
You are running on the same hardware, so should be drop-in replacement for your SW setup.

An image would be really cool, maybe we can track down the 0.2 fps I’m missing on my GX4 with that method as well.

Several Headland HT## chipset based 286 motherboards passed through here.
One of them (HT12P-16A) shown really good performance comparable to VLSI-200, so i was curious if the later HT18c variant will bring a surprise.
Headland HT18c is a hybrid 286-386SX silicon and one of the latest chipsets supporting 286 processors.
Finally managed to obtain a board based on it.

Protech PM286

It came in excellent condition with 16MHz rated CPU. Replaced it with 25MHz Harris one.
Inserted fast Cyrix FPU.
Soldered crystal oscillator socket.
BIOS is 386-like. Does not offer any options for memory timings.
Jumpers allow combining of on-board DIP and SIMM memory. Gives +1Mb RAM. Turned it off. Relied on trusted SIMM memory modules only.
There is jumper for FPU speed, but nothing for memory wait states.

Board fired up right away at 25MHz with up to 16Mb RAM (4x4Mb 50ns).
Few short iterations later it was ticking reliably at 31MHz (62MHz oscillator).
This required active CPU air cooling and 4x1Mb 40ns rated memory modules. Glad i purchased some from a fellow member here.
Can work at 35MHz, but system is unstable.

There are 2 empty places for additional crystal oscillators.
Spent a moment to socket them.
Turns out J9 can switch between the primary one (on picture) and the one next to it.
J1 in position 1-2 will feed FPU with signal from the primary oscillator. J1 in position 2-3 will attempt to use signal from the oscillator next to the FPU if available, otherwise FPU will not work. If that oscillator is faster than 30MHz the FPU starts hanging the system.
I am not 100% sure what exactly its role is because FPU works with any frequencies between 16.143 and 30 MHz, but performance does not change which indicates some sort of synchronization functionality only.
J2 - does seemingly nothing - no observable difference if it is open or closed.

Not much more to add.
Long story short is - 31MHz is the peak. Stuff just works.
So, let's look at what this means.

The usual stats:

Not bad, not bad at all.

Just a reminder of what the Autodesk Chaos test is.
Offline computation to produce an image based on the Mandelbrot fractal, MSUN preset.

benchmark results

Not the fastest clock-to-clock chipset. Slower at interactive graphics.
Otherwise at that CPU/FPU speed it offers the fastest 286 offline compute that i have seen so far !

Cool, is the AT bus nailed to half the 14.xxx mhz reference clock or something? 7-ish Mhz?

No idea about the AT bus. Don't have oscilloscope to check.
There is a place for second oscillator. Hmm.

@Chadti99, will pass you the partition images tomorrow.

Chadti99 wrote on 2022-01-02, 03:33:

pshipkov wrote on 2022-01-01, 20:17:

I am looking at the same manual. Both 64kx8 and 128kx8 work. Wanted to check if either one worked better for you, etc. […]
Show full quote

I am looking at the same manual. Both 64kx8 and 128kx8 work. Wanted to check if either one worked better for you, etc.

Back to LSD.
Stared a bit longer at your last notes and screenshots, compared them to my previous posts and ran some brief tests here.
1. The 20.4 fps in Quake 1 was achieved with CL-5480 graphics card. That's 0.1 fps higher than what i get with Voodoo3 and MGA (which is 0.3 fps).
2. There is one difference between our BIOS setups - i set DRAM SPEED to FASTER which allows cache timings to be 2-1-2. This is better for performance. That is with 50ns EDO RAM - the same module you use.

Thanks for reviewing this again, I’m now running 212 cache settings on the LS-D and hitting the same 20.3fps at 180MHz. Changing the DRAM SPEED to FASTER was the key to lowering cache timings!

I could of sworn I was Windows stable at cache settings of 212 on my LSD, but now I can’t seem to complete a GLQuake Timedemo. My bios settings reset at some point while messing around with the the variable resistor. Maybe I’ve missed something, quite frustrating.

We know that story very well.
Sucks when happens.

Probably cpu cooling. Give it 9-12v Peltier and high chance it passes.

novanick wrote on 2022-01-22, 04:20:

Is there any chance I can get you to send me a digital copy of this motherboard manual. Last year I came across this board and had to do some repairs, just getting ready to set it up and wouldn't mind giving the manual a once over.

Thank you very much,

Attached you the pictures buyer forwarded.

pshipkov wrote on 2022-01-27, 04:41:

We know that story very well.
It socks when it happens.

I bet it is cpu cooling. Give it 9-12v Peltier and high chance it passes.

Yah that is the frustrating part, it’s at 180MHz(60x3) which hasn’t required peltier on this specific chip. If I drop cache timings back to 312/fastest I’m fine. I should really start taking better notes but I’m quite certain 212/Faster was at least GLQuake stable at one point for me.

Back to the variable resistor, is it not possible to measure voltage while in use? I think maybe that was my issue. If I remove the CPU I seem to able to adjust and get good measurements.

I thought you arecrunning 180 on 5v Peltier, that's why suggested to increase the voltage.
But if you were running it naked, then 5v would be ok.

pshipkov wrote on 2022-01-30, 00:01:

I thought you arecrunning 180 on 5v Peltier, that's why suggested to increase the voltage.
But if you were running it naked, then 5v would be ok.

Cooling the CPU will help stability with lower cache timings? Maybe I was using some thermal paste before.

It’s so crazy because now 212/Faster isn’t even dos stable now, not really sure what changed.

Yes, I noticed that cooling the cpu affects cache and mem timings in OC situations like yours.
But it can be simply a component just starting to show its age..

Does it work with Peltier on top?

pshipkov wrote on 2022-01-30, 01:01:

Yes, I noticed that cooling the cpu affects cache and mem timings in OC situations like yours.
But it can be simply a component just starting to show its age..

Does it work with Peltier on top?

I’ll give it a shot. Could def be aging components.

In other news, has anyone discussed something like this to retro faster cache modules on boards with dip sockets? Spotted this browsing eBay.

Havent seen these as a commercial product.
I guess their market will be very small because there are 10/12ns rated DIP chips available that are more than enough for 486 and Pentium 1 class hardware.

I sort of recall someone doing this or at least prototyping it some years ago, but forget who.

Are those 15 ns SOJ on a DIP adapter? If so, what's the benefit? Also, won't the leads on the PCB be a bit wide and stretch out the motherboard's DIP sockets?

Almost an year ago posted here about AbitAB-PB4 revision 1.3 486 class motherboard.
Great hardware for up to 160MHz, but couple of things bothered me with it:
- the on-board clock generator supports 60/66MHz frequencies required for 180/200MHz to CPU but there is no jumper configuration to enable them. Manual talks about JP31 but it is not present in PCB revision 1.3, so FSB cannot be higher than 40MHz.
- completely unstable system with P24T (POD83/100)

I happen to have revision 1.5 of the same motherboard but it was stuck at 5V to CPU which prevented me from running most 4x86/5x86 processors on it.
Been on my to-do list to fix that and examine it properly.

Most notes from the original post hold true for this one too, with the next differences:
The non-standard slot from revision 1.3 is replaced with normal 16-bit ISA slot.
JP31 is there.
JP12 is not soldered in position 1-2.

Comparing the voltage regulators of the two boards (upper-right corner) gives a hint about the possible problem.

revision 1.3:

revision 1.5:

Swapped the VRMs - instant success with 1.5.
1.3 got stuck at 5V. Will have to upgrade it to the same 5 legged linear VRM at some point later.
Clear sign of design flaw in both PCBs.

--- Am5x86 @ 160MHz

With proper VRM everything worked as expected with 4x86 and 5x86 CPUs.
With all BIOS settings on max the system gets quite picky about RAM modules and L2 cache chips.
Took me a moment to find the right combination.
Ended up with 256Kb L2 cache and 64Mb 60ns EDO RAM.
No point to go past 64Mb since it will be out of the cacheable range.

SpeedSys looks the same as the screenshot from the provided above link to the first post.

benchmark results

It is clear that Abit made optimizations to revision 1.5 as it is consistently faster than 1.3 with WinTune2 (accelerated Windows GUI) being the only exception. That difference is more of a rounding error than anything else.
The improvements get performance closer to the best numbers on record. Pretty nice.
IDE interface without driver hits around 5Mb/s reported by SpeedSys and CoreTest.
With driver the local storage i/o hits 9-10Mb/s which is on par with Asus PVI (best on-board IDE, etc.)

--- Am5x86 @ 180MHz

No combination of jumpers seems to produce 60MHz FSB. Will check this more carefully in the coming days, but so far no lights beyond 50MHz FSB.

--- Am5x86 @200MHz

3x66MHz is not possible. No jumper settings seem to produce 66MHz FSB.

System is fully stable at 4x50MHz, 200MHz CPU.
All BIOS settings on max except:
DRAM READ TIMING = FAST (best is FASTEST)
SRAM READ TIMING = 3-1-1-1 (best is 2-1-1-1)
4V to CPU for complete stability.

SpeedSys reports really nice IDE metrics.

benchmark results

Obviously not the best performer out there, but ticks pretty well for what it is.
Feels really solid - completely trouble free - very satisfying.

--- P24T (POD83/100)

PODs work much better on revision 1.5 than 1.3 but still very flaky.
No amount of tweaking produced stable system - different voltages, most conservative BIOS settings, etc.
Really disappointing.

feipoa wrote on 2022-01-30, 21:15:

I sort of recall someone doing this or at least prototyping it some years ago, but forget who.

Are those 15 ns SOJ on a DIP adapter? If so, what's the benefit? Also, won't the leads on the PCB be a bit wide and stretch out the motherboard's DIP sockets?

They blurry pixels seem to show 15ns chips.
The only interesting configuration I can think of is 32-pin pcb for 64k x 8 chips since faster than 15ns ones are in short supply.

“System is fully stable at 4x50MHz, 200MHz CPU.”

@pshipkov, is this with peltier?