OK, it may also depend on some impurities in silicon during manufacturing process how the resulting gate leakage would be, so I would expect high variance of t_refresh at given temp, probably need to be figured by trial-error. On some of my 386/486 MBs there is option for various refresh timings but not in range of hundreds us, if I remember well max. is about 50-60us. I didn't test how it affect the system performance but I though L1/2 cache should help to make the difference small but feipoa's results show greater differences than I'd expect...

The 24% improvement you are seeing is because of no hidden refresh and using BARB to invalidate the L1 cache every hold state. Probably if you are using a regular 486DX, you would not see such a large performance increase when using DRAM.COM. I think up to 5% in cases not using BARB.

I recall reading in some literature that older RAM from the 80's needed more frequent refreshes compared to more modern RAM. I have a late era 386 board w/VLB that has BIOS refresh options from 15-120 uS.

Aha, I enable hidded refresh when available...

Was able to get 96MHz (2 x 48, using Migron's clockgen) ticking somewhat reliably today.
DTK PEM-0030Y, 256Kb L2 cache, 16Mb RAM, Cirrus Logic GD-5434, generic IDE controller, ~3.9V to CPU, poor air cooling (this processor really does not care about temperatures at all).
All BIOS settings on max except DRAM WAIT STATE PARAMTER = 1 (best is 0).
The CPU can clearly do more, but some of the other components are starting to hit their limit.
Will have to swap them with better ones before continuing. At minimum i can do 100MHz runs but with inflated wait states = lower performance.

Run DOS interactive graphic tests only to feel things out. Will try to reach clean 100MHz before spend time on stability and the more interesting tests.
Wolf3D: 87.9 fps <- impressive, only Alaris Cougar II with Ark1000VL at 3x33 does better
Superscape: 33.3 fps
PC Player: 8.4 fps
Doom: 20.34 fps <- not sure why so low, relatively
Quake 1: 5.4 fps <- impressive, best score so far, hope it is not a skewed clock at play

Couple of screenshots and photos.
Notice the visual artefacts. It is the motherboard. It is caused by a combination of high-enough base frequency, some VGA chips and tight wait states.
Good enough for testing, it does not affect stability or anything. Should have switch to DTK PEM-4036Y for the photo session, but too lazy for it right now.

Feipoa, notice how at 96MHz TIMING PARAMETER SELECTION can be set to NORMAL and DRAM WAIT STATE SELECTION can be set to 1.
I think i entered a sweet spot of sort because these parameters cannot be this tight for anything above 40MHz FSB.

I am still not confident this can be a fully stable system. For now, exploring where the upper bound is.
Good chance i have to step-down below the 90MHz mark to get it stable.

There are actually "sweet spots" in circuits, where the noise is tuned to a minimum, so you might be right.

Think how buildings are made to deal with earthquakes with mass tuned dampers.
The damper is chosen to match the resonant frequency of the building vs earthquake wavelengths.

It's not the same same in circuits, but to a large extent, a wave is a wave, and acts like a wave.

These results are certainly inspiring! Unfortunately, I don't have a loose 4030Y or 4036Y for benchtop testing.

At 80-90 Mhz, I've always been able to run DRAM at 1 ws and timing at Normal. I find it curious that you've found otherwise.

Some questions:

1) Did you need to drop from 32 MB to 16 MB to get 96 MHz prematurely DOS stable?

2) How are you running voltage to the CPU? Are you first going through the buck, then re-regulating with the linear MIC?

3) I noticed you running ISA at max speed. Did reducing this eliminate the artefacting? Have you found that increasing ISA speed beyond some value doesn't necessarily improve DOOM benchmark results?

4) Do you just have a heatsink sitting on the bare CPU ceramic, or is there heatsink compound on there as well?

5) Crystal size on that migron PCB is huge. Did you find that the larger crystal helped with respect to overclocking compared to the smaller crystal? What crystal value are you using?

6) Was it only 1 CPU out of 6 tested could do 90+ MHz?

Interesting that you can run at 40+ mhz with these two parameters set to NORMAL and 1. I know this leads to instabilities in some of the more serious tests for me.

1. 32mb is fine. The other 4x4mb are sitting on the bench. It takes trusted memory modules to get to 90+ mhz with tight timings.
2. Buck only. Bypassing all the gear on the adapter. If 100mhz turns out well i will be removing pieces from the adapter.
3. If isa bus wait states get decreased or base frequency gets decreased for this mobo the artefacts disappear. Not able to run the isa bus at min wait states is a direct disqualification for the 386 mobos here. So far faster isa bus results in ever increasing overall performance.
4. No heatsink paste. The heatsink kind of does nothing really. These cpus are very cold even under heavy load.
5. There are two types of the 2-pin crystals. The only 24mhz one i had is with the tall body. So far the only case where the crystal mattered is to get BL3 cpus running at 110mhz stably with msi ms-3124, nowhere else.
6. Only one of the cpus turned out well. Actually the other five appear dead as of right now. Dont ask me how i got to this point.

Are benchmark scores faster without the FPU?

Interesting. I have noticed an asymptotic limit with increasing ISA speed and DOOM scores. Perhaps you are witnessing this with your DOOM score, considering it is lower than you expected.

Do we know the drop-out on the buck? Since you are only running at 3.9 V, I bet you could just wire in the buck to Vin/Vout on the PCB. In this case, you'd run Vin as 5 V to the buck and not need to selectively remove all the Vcc pins on the interposer in order to run Vin as 12V.

I need to get more of these interposers assembled for use with buck only, but I'm out of the PGA pins. I ordered some recently but the gold plating was so thin that I was questioning whether they were actually plated. I'll need to try another supplier, but I was questioning if all the supply is like this now.

If the heatsink does nearly nothing, are you able to run it for an hour without the fan on the heatsink and still survive Quake? Without the fan, I find the heatsink gets pretty darn hot after an hour of testing.

@myne
Yes, such "resonance" is observed in some cases.

@MikeSG
FPU is only a plus - for tests that leverage it.
But FPUs are often problematic for higher overclocks.

@feipoa
The relatively low doom score is a bit puzzling, that's why marked it as such. It should be 24 or more fps.
But that is unrelated to ISA bus perf plateauing. It is just low in general for some reason. Probably missing something since all other results are just great.

Yes, I can switch back to adapter native voltage control, now that i have a magical CPU, but if i get the itch at some point later to play with SXL2 chips it will be beneficial to have the room the buckreg provides.
Cannot remember if i mentioned it already, but my ATX PSU outputs 4.8-4.9 volts only on the 5V rail. I don't like that. It was a factor for the more flaky SXL2 processors, that's why moved to the buckreg.
Recently Synology NAS here died. Its PSU outputs 5.2V, so thinking to start using it going forward, but it does not do -12V for some reason, so need to mod it first. Maybe with it on the test bench i won't need the buckreg anymore. Anyhow.

I can leave Quake 1 to roll but didn't do long hours at all.

As for the new pins - you are not thinking right about them. White gold is the trend in high fashion now. Feel good.

You will either create -12V from +12V or from -5V? Reminds me of some ATX to AT adaptors which have added this conversion to the cable. Is it easier to buy an old AT PSU with -12V already in place? My best units will output ~5.15 V, which after the connector, and traversing the motherboard and CPU pins might be around 5.10 V.

According to the supplier of the pins, the gold layer is "very thin". For me, it isn't worth the effort of making another interposer containing gold coating which may blow off. I'll need to order from elsewhere and try my luck.

Why do you suppose the buck shows initial promise at 90 MHz while the linear does not? Is the 5V line going into the linear too noisy compared to the bare 12 V dangling from the PSU? If you run the MIC at 4.1 V, what is the failure mode at 90 MHz?

These TI486SXL2-66 chips with factory heatsinks - they are using a thermal pad that must be a few decades old. Does anyone feel that the thermal transfer proprieties of the pad's adhesive will have sufficiently diminished over this time?

Will derive -12 from +12.
The ATX to AT adapter i have is of most simplistic design.
The Synology ATX PSU is very compact, like 1/4th to 1/3rd of a standard PSU dimensions. That's why i am considering it.

Buck vs linear voltage regulators
Initially i didn't have much luck with the on-board LVR, not sure why. Now i bypassed it and having much more luck with the buckreg. But there is also a chance i was doing something wrong before. Back then i was distracted with other things which increase the chance of user errors. Will verify these things soon, now when i am at a place that allows me to build from.
But what i stated in my previous reply was that i am leaning towards keeping the buck because it makes me +5V PSU rail independent. But this may not be a factor going forward, so this initial thought may be revisited later and i will conform to t he onboard LVR.

No idea about how the aged material degrades heat propagation. These chips are cold, so thermal conductivity of the pad is probably a no factor.

---

actually the Synology PSU is more like 4 times smaller. it may appear longer that the standart ATX PSU, but that is an optical illusion from the perspective and lens distortion.

I am looking forward to your stability testing at 90 or 96 MHz and resolving the mystery/uncertainty around linear vs. switching.

I noticed that you still have the MIC linear regulator installed on the interposer in your previous photo, yet are using the buck on the fan's 5V header. I assume you are merely cutting the MIC's enable lead to disable it?

Buck/Linear regulators:
The response characteristics of the Buck Regulator are far better than Linear, as they can respond to voltage dips depending on switching speed. If this is working better than Linear Regulators, it shows a lack a capacitance absorbing the dips..? DX4-ODP/R, and POD linear regulators are accompanied by 3x 3.3UF instead of one 10UF for this...

TI486SXL2-66 thermal glue:
486 CPUs in general take time to heat up, they don't reach 100C in one second like modern CPUs. So even modern thermal pads can transfer heat enough...

FPU / SXL2(C) systems:
On my 16-bit SXL2(C) 386sx system, removing the FPU gave 1 FPS in 3DBench. So I was wondering whether it made a difference on the DX boards at high frequency.

For interest / comparison, these are my top three benchmarks in 3DBench:
First: 31.2 @ 100Mhz (50Mhz sys), 20Mhz ISA clock, 1 CAS Extend WS (read), no FPU. No BARB, flush/memw line not connected.
Second: 30.3 @ 90Mhz (45Mhz sys), 22.5Mhz ISA clock, 1 CAS Extend WS (read), no FPU. No BARB, flush/memw line not connected.
Third: 30.3 @ 50Mhz (50Mhz sys), 20Mhz ISA clock, 1 CAS Extend WS (read), no FPU. No BARB, flush/memw line not connected.

The sys clock at 50Mhz is the optimum zone but the keyboard controller stops working so I'm working through that.

Has anyone tried 50Mhz sys clock, but not clock doubling?

MikeSG wrote on 2025-01-12, 06:39:

Buck/Linear regulators:
The response characteristics of the Buck Regulator are far better than Linear, as they can respond to voltage dips depending on switching speed. If this is working better than Linear Regulators, it shows a lack a capacitance absorbing the dips..? DX4-ODP/R, and POD linear regulators are accompanied by 3x 3.3UF instead of one 10UF for this...

You are right that buckregs are better than linear ones, but shouldn't have an impact on what we do in this case in terms of voltage dips and so on. Or, do you see it differently?

MikeSG wrote on 2025-01-12, 06:39:

On my 16-bit SXL2(C) 386sx system, removing the FPU gave 1 FPS in 3DBench. So I was wondering whether it made a difference on the DX boards at high frequency.

This is interesting. Never saw anything like that with 386DX stuff.

MikeSG wrote on 2025-01-12, 06:39:

For interest / comparison, these are my top three benchmarks in 3DBench: First: 31.2 @ 100Mhz (50Mhz sys), 20Mhz ISA clock, 1 CA […]
Show full quote

For interest / comparison, these are my top three benchmarks in 3DBench:
First: 31.2 @ 100Mhz (50Mhz sys), 20Mhz ISA clock, 1 CAS Extend WS (read), no FPU. No BARB, flush/memw line not connected.
Second: 30.3 @ 90Mhz (45Mhz sys), 22.5Mhz ISA clock, 1 CAS Extend WS (read), no FPU. No BARB, flush/memw line not connected.
Third: 30.3 @ 50Mhz (50Mhz sys), 20Mhz ISA clock, 1 CAS Extend WS (read), no FPU. No BARB, flush/memw line not connected.

The sys clock at 50Mhz is the optimum zone but the keyboard controller stops working so I'm working through that.

Has anyone tried 50Mhz sys clock, but not clock doubling?

So, you are running 2x50MHz SXL2 over there?
Nice. Stably? At what voltage?
I don't remember reading early in the thread about it.

As for native 1x50MHz - yes - that's very easy. ISA bus at 22.5MHz, etc.
I had a post about keyboard controllers here. They matter. Take a look.

feipoa wrote on 2025-01-12, 03:07:

I noticed that you still have the MIC linear regulator installed on the interposer in your previous photo, yet are using the buck on the fan's 5V header. I assume you are merely cutting the MIC's enable lead to disable it?

This is what i got for one of the adapters.

Thanks for the photo of the interposer bottom. What does connecting the LDO's Vin pin to MIO do? Are you grounding the other end of the MIO pin to the PSU to disable the LDO?

EDIT: I think MikeSG mentioned earlier that he is running the SXLC2, so 16-bit only. Overclocking success with 16-bit probably doesn't directly translate to 32-bit (SXL2) success. Nonetheless, I'd be interested if he was able to run Windows 3.1 at 100 MHz using an SXLC2.

Yes. Buck OUT to LDO IN OUT. Other MIO pin to PSU GND. This was the least intrusive way i could think of. Do you see a problem with it?

Yes, i remember long ago checking the performance progression in DOS real mode between 286, 386DX, and 486DX processors. L1/L2 caches disabled for 386 and 486.
Motherboards and chipsets matter. Won't go into detail, but let's say i put effort to get closer to apples-to-apples comparison.
In short - perf progression is fairly linear and directly correlated to the base frequency.

Observed something similar from my experience with 286-to-386 upgrades where some of the adapters came with SXLC2 CPUs. Their perf characteristics are not that different than the SXL2 cousins. Something that we see from Mike's notes above. Superscape results from 96MHz SXL2 is basically the same as 100MHz SXLC2.
In addition, SXLC2 and SXL2 seem to be fairly similar in terms of stability and overclocking.

Did further testing with DTK PEM-4036Y. Anticipated improved stability compared to DTK PEM-0030Y but that was not the case.
Along the way i noticed something interesting - PEM-0030Y seems to be the better motherboard for SXL2 chips compared to 4036Y/YB.
At 40MHz FSB and up 0030Y can handle TIMING PARAMETER SELECTION = NORMAL and DRAM WAIT STATE SELECTION = 1 compared to EXTENDED and 2 for the later 4036 variants.
So back to 0030Y.
Next up - slap a Peltier on the CPU. I did this before with the more flaky SXL2 CPUs. It was not a factor but will try once more with this very promising chip - just in case.

pshipkov wrote on 2025-01-13, 00:46:

Yes. Buck OUT to LDO IN. Other MIO pin to PSU GND.

I think you meant BUCK OUT to LDO OUT. If it was Buck OUT to LDO IN (grounded via PSU), you'd have a short.

pshipkov wrote on 2025-01-13, 00:46:

This was the least intrusive way i could think of. Do you see a problem with it?

No problem with it per se, I was more or less just trying to figure out what you were doing and it was not apparent from the first photo you provided. I'm not sure if having the LDO in place, with its Vout tried to the buck Vout, could cause any undesirable results, even though it is set to disabled. I recall there being some discussion in this regard concerning connecting the FLT# to GND on Am386DX QFP CPUs. Seemed that removing the Am386DX, rather than disabling it, was preferred. However, as you are having some initial success, maybe it's fine?

pshipkov wrote on 2025-01-13, 00:46:

Yes, i remember long ago checking the performance progression in DOS real mode between 286, 386DX, and 486DX processors. L1/L2 […]
Show full quote

Yes, i remember long ago checking the performance progression in DOS real mode between 286, 386DX, and 486DX processors. L1/L2 caches disabled for 386 and 486.
Motherboards and chipsets matter. Won't go into detail, but let's say i put effort to get closer to apples-to-apples comparison.
In short - perf progression is fairly linear and directly correlated to the base frequency.

Observed something similar from my experience with 286-to-386 upgrades where some of the adapters came with SXLC2 CPUs. Their perf characteristics are not that different than the SXL2 cousins. Something that we see from Mike's notes above. Superscape results from 96MHz SXL2 is basically the same as 100MHz SXLC2.
In addition, SXLC2 and SXL2 seem to be fairly similar in terms of stability and overclocking.

The intention of my comment was to say that an SXLC2-66 should overclock a bit better compared to an SXL2-66 because half the data bus is utilised. Wouldn't sending 16-bits per clock be less "stressful" on the CPU compared to 32-bits, especially when you are exceeding the manufacturer's limits? Wouldn't it also produce less heat, further aiding overclock-ability? I vaguely recall mkarcher or one of the other Vogons gurus discussing this at some point.

pshipkov wrote on 2025-01-13, 00:46:

In addition, SXLC2 and SXL2 seem to be fairly similar in terms of stability and overclocking.

Where did you find an SXLC2-66 QFP to run this comparison? Could you point me to your results? I probably missed this at some point.

pshipkov wrote on 2025-01-13, 01:17:

Did further testing with DTK PEM-4036Y. Anticipated improved stability compared to DTK PEM-0030Y but that was not the case. Alon […]
Show full quote

Did further testing with DTK PEM-4036Y. Anticipated improved stability compared to DTK PEM-0030Y but that was not the case.
Along the way i noticed something interesting - PEM-0030Y seems to be the better motherboard for SXL2 chips compared to 4036Y/YB.
At 40MHz FSB and up 0030Y can handle TIMING PARAMETER SELECTION = NORMAL and DRAM WAIT STATE SELECTION = 1 compared to EXTENDED and 2 for the later 4036 variants.
So back to 0030Y.
Next up - slap a Peltier on the CPU. I did this before with the more flaky SXL2 CPUs. It was not a factor but will try once more with this very promising chip - just in case.

This explains why I've always been able to run timing at 1ws and Normal - I have the 4030Y and not the 4036Y. But also on that noname HaydnII board, I can use 1 ws and normal.