Wtih 32 MB installed, 3.9 fps is also what I get in Quake 1.06 at 90 MHz, but only with 11.25 MHz ISA.

It is good to know that the peltiers helped nil. However, I did notice some thermal factor involved because if I removed my fan and ran the unit with just the heatsink at 80+ MHz, it would start throwing Windows errors. Also, I could run 90 MHz in the winter stable at 5.25 V, but not in the summer. I've had to take it down to 87.4 MHz and 5.05 V until I can investigate this further.

You noted that the BL3 at 110 MHz on the same Symphony motherboard yielded 28.3 fps, yet the SXL2-90 (22.5 MHz ISA) yielded 34 fps. What ISA frequency was that BL3-110 running at?

If you take the SXL2 down to 80 MHz and run the ISA at 10, 13, 17, and 20 MHz, what DOOM scores do you receive? Is the system stable?

Peering thru my notes, I have the following written down:

SXL2 at 80 MHz
10.0 MHz ISA = 3607 = 20.71 fps
13.3 MHz ISA = 3396 = 21.99 fps
16.0 MHz ISA = 3268 = 22.86 fps
20 MHz ISA = 3164 = 23.61 fps

From 10 to 20 MHz ISA, I'm noticing a 14% improvement.

At 90 MHz, you noted DOOM = 23.97 fps at 11.25 MHz and 33.9 fps at 22.5 MHz. This is a 41% improvement just by doubling the ISA speed, but it is far from the 14% I'm seeing at 80 MHz. Are you able to verify at 80 MHz?

I also have in my notes that at 20 MHz ISA, Windows 3.11 could not load, and that was without my NIC installed. Notes indicate that max ISA my NIC can run at is 13.3 MHz.

Anyone tried the Floating Point Ready Return option in BIOS? It's in the BIOS modding tools if not in the BIOS itself.

Option 1 - FPU returns READY. Option 2 - Chipset determines FPU READY. (Wording might not be exact.)

May be a little more speed from the FPU in Quake.

Mistake. It's chipset dependant for another board. C&T 82C836 386sx

Quake 1 is CPU/FPU heavy. Video system just sits on its hands most of the time, so ISA bus frequency is the minor factor.
It takes fast Pentium or later processor to turn it around.

I remember you saying recently that you hit this hurdle with the warmer environment temperature.

ISA bus runs at 27.5MHz when running the BL3 CPUs at 110MHz (2x55).

System is not stable at 80MHz with none of the 6 CPUs here. Regardless of other factors. It is either the CPUs or something is just wrong with the two adapters i assembled. Starting to suspect that actually. I don't have the 6 big capacitors on the top installed. May be something like that.

Will run a quick test at 80MHz and the ISA bus frequencies you listed, to cross-check numbers.

The performance difference between 11.25 and 22.5 appears staggering. And that's why it is not real. Here is the proof:
This video shows 2 Wolf3D test runs side by side, both captured with my phone at 30 fps and synchronized in post to the first frame of movement.
Btw, ignore the artefacts on screen. I had a flaky video ram chip in the stash. Pulled the shortest straw early today while adding a second megabyte of RAM to the STB Nitro card. Too lazy to replace right now.

On the left side we have ISA bus at 11.25MHz. Wolf3D reports 57.3 fps. In the video the "results" screen appears 12.18 seconds in, which is roughly in line with 712 frames divided by 12.18 seconds = 58 fps. Close enough.
On the right side ISA bus is at 22.5MHz. Wolf3D reports 128.4 fps. In the video the "results" screen appears 9.22 seconds in. So, 712 frames divided by 9.22 seconds = 77 fps. That is not 128 fps. The time warp is 44%

I bet the inflated results from the other tests are incorrect too. For example, scaling Doom's 34 fps down by ~40% brings it to ~23 fps, which is a lot more plausible.
Will verify that tomorrow using the above approach.

@MikeSG
I didn't know about that until your note. Added it to my to-do list.
One more note from me along the same lines - Quake 1 likes ULSI FPUs better than Cyrix FasMath.

pshipkov wrote on 2024-07-29, 07:40:

System is not stable at 80MHz with none of the 6 CPUs here. Regardless of other factors. It is either the CPUs or something is just wrong with the two adapters i assembled. Starting to suspect that actually. I don't have the 6 big capacitors on the top installed. May be something like that.

This would give me pause. Either you are very very unlucky, or we are missing something. Of the 13 CPUs I tested, only 2 weren't able to handle 80 MHz reliably, and the failing 2 could do 75 MHz w/fan. You shouldn't need the larger caps, but I had them installed originally for testing purposes.

I'm not sure what else could be at fault with your adaptor, other than soldering or the VRM. The primary differences between our interposers are:
1) male PGA pins flipped (shouldn't matter for what we are doing, but your orientation may eventually stretch out the female PGA holes).
2) different VRM brand and model. See note after this list.
3) different assembler/solder. It's a long shot, but maybe add more solder to your VRM's Vin and Vout and increase dwell time for the heat.
4) different capacitor quantity. You could try populating the pads, but they are only for Vin on Alpha 2, thus of little value to what the CPU sees.
5) different PCB. I'm primarily using Alpha 1 and have used it exclusively for 90 MHz. You are using Alpha 2. Maybe assemble Alpha 2 if all else fails.
6) different heatsink/fan, but shouldn't matter. I'm only giving 10 cfm to my factory heatsink.
7) EDIT, NEW THOUGHT: When I was doing testing and assembly, I noticed that sometimes you can get a solder blob between the PGA168 pins on the interposer. They can be difficult to see and may look harmless, but I had one case in which a small blob caused my interposer to be flakey. Similarly, I had one of the 0805 SMD caps soldered on, which looked perfectly fine, but caused flakey performance. I removed it, and all was good.

Now that I think about it, I'm pretty sure I had some marginal MIC VRM from an unknown source during the COVID silicon shortage. While all of the MIC regulators needed more voltage than the LP's, there was one MIC I had to discard because the system was unstable with it, at any voltage. I don't recall the source of that MIC. Unfortunately, the LP doesn't have a mounting hole for the heatsink, so for what you are doing now (greater than 4.9 V), you'd want to add an in-line switching regulator so that your VRM doesn't need a heatsink, or glue on a heatsink to the LP regulator.

pshipkov wrote on 2024-07-29, 07:40:

On the left side we have ISA bus at 11.25MHz. Wolf3D reports 57.3 fps. In the video the "results" screen appears 12.18 seconds i […]
Show full quote

On the left side we have ISA bus at 11.25MHz. Wolf3D reports 57.3 fps. In the video the "results" screen appears 12.18 seconds in, which is roughly in line with 712 frames divided by 12.18 seconds = 58 fps. Close enough.
On the right side ISA bus is at 22.5MHz. Wolf3D reports 128.4 fps. In the video the "results" screen appears 9.22 seconds in. So, 712 frames divided by 9.22 seconds = 77 fps. That is not 128 fps. The time warp is 44%

I bet the inflated results from the other tests are incorrect too. For example, scaling Doom's 34 fps down by ~40% brings it to ~23 fps, which is a lot more plausible.
Will verify that tomorrow using the above approach.

Ahh, nice work. Now we have some correction. Looking forward to the fixed DOOM scores and your 80 MHz comparison numbers.

MikeSG wrote on 2024-07-29, 07:01:

Anyone tried the Floating Point Ready Return option in BIOS? It's in the BIOS modding tools if not in the BIOS itself.

Option 1 - FPU returns READY. Option 2 - Chipset determines FPU READY. (Wording might not be exact.)

May be a little more speed from the FPU in Quake.

I haven't tried this. Which BIOS modding tool? I do not recall seeing this in any BIOS.

BIOS modding tool:
AMISetup v2.99.0. "Coprocessor READY cntrl by". 0 = FPU. 1 = chipset.
These options are chipset dependant.... so may not be there. It's there for the C&T 82C836 386sx. There are 6x SYM (symphony) 1991 chipsets listed as compatible for the tool.
It's chipset dependant for the C&T 82C836 386sx, my mistake...

Voltage:
Another thing to think about is XTAL High voltage v Cpu VCC. It's not High at 2V like other pins.

In the TI486SXL Microprocessors Reference Guide, Page 210, Electrical Specifications 5-5

"CLK2 High-level input voltage: Min VCC-0.3 ...Max VCC5+ 0.3 V"

If CPU VCC is 5.0, minimum motherboard VCC (XTAL VCC) would be 4.7v.

MikeSG wrote on 2024-07-29, 10:19:

BIOS modding tool:
AMISetup v2.99.0. "Coprocessor READY cntrl by". 0 = FPU. 1 = chipset.
These options are chipset dependant.... so may not be there. It's there for the C&T 82C836 386sx. There are 6x SYM (symphony) 1991 chipsets listed as compatible for the tool.

I don't recall seeing this option in AMISETUP, but I'll look again when I have my toys out in October.

MikeSG wrote on 2024-07-29, 10:19:

In the TI486SXL Microprocessors Reference Guide, Page 210, Electrical Specifications 5-5

"CLK2 High-level input voltage: Min VCC-0.3 ...Max VCC5+ 0.3 V"

If CPU VCC is 5.0, minimum motherboard VCC (XTAL VCC) would be 4.7v.

I remember you quoting this in the past and for my setups, I've always been sure to fall within this range. For my PSU, I get 5.08 V at the cyrstal oscillator, thus the max I've taken CPU Vcc to was 5.35 V. For pshipkov's setup, he is using an ATX PSU, and after the AT adaptor, I think he was noting only about 4.7 or 4.8 V to the motherboard. Thus, you could argue that since he is running the CPU at 5.3 V, he has exceeded that TI specification. However, he was also not able to achieve stability at 80 MHz and 4.5 V to the CPU on all 7 of his SXL2-66 chips, which is strange. Thus, I feel the root cause of the issue is still undiscovered.

One thing that can be happening here is high ISA clock bugging 8254 programmable interval timers source of truth (14.31818 MHz). I dont know of any testing software to measure real PIT clock, hmm maybe just observing if dos Time goes bad after a while ?

I have the same idea that the timer clock source used by games for timing FPS may be affected by unusual high ISA clock. Maybe it's specific for that chipset. Maybe also other systems with unusual ISA speed benchmark results may be fooled. It would be possible to look at doom game sources how timer is used and make some small test utility that would compare this timer speed against CMOS RTC to see if the system has timer precise enough... good to know...

About voltage regulators - may depend on your part supplier, you may expect anything from china, a recycled or partially damaged component or different component that was sanded and remarked. For some comparison you could measure Vpp noise level on the output if there would be some significant diffs between varoius regs...

I have noticed on certain motherboards that increasing the ISA frequency would actually show reduced benchmark scores. This is the first I've heard of this occurrence in the other direction and by this magnitude.

I do not recall where my faulty MIC regulator originated. I'm pretty sure it was from digikey (I had 2 from them in 2017), but could have eventually been mixed in with the units from eBay Taiwan or eBay China. For all the MIC regulators received, they did need 0.15 - 0.2 V more than the LP regulators for stable operation. Then there was the one MIC that wouldn't run stable at any voltage which I discarded.

The issue happens only with SXL2-66 CPU in adapter.
PGA/PQF132 chips are fine 386DX, SXL2-50, BL3 running at various speeds, including 50MHz base, are all fine.

Btw, the MIC regulator on the 12V modified adapter died yesterday.

lol, oh no! What sustained temperatures were you recording on the MIC regulator? At this point, I'd bypass the linear regulator entirely and wire in a $1 switching regulator module. Will PM you a link.

Just for curiosity, maybe you already know/seen it, I didn't before... one user on a local forum sent a photo of his hybrid 386/486MB. 1st I guess he's talking about 486DLC but no, his MB has full-featured 486DX socket with DX2 plugged inside:
http://www.ebastlirna.cz/modules/Forums/files/doska_290.jpg
just a pity it's missing a VLB slot for faster graphics but I found that even such MBs (with Opti chipset) existed...
I guess that performance would be close to a regular 486 system, that it just enabled a cheap starter onboard 386 CPU for users that could't pay for true 486 during upgrade letting them to upgrade later...

Looking over at the Toshiba T5200 thread they have a list of PGA132 486DX CPUs that will work with that device.
Re: Toshiba T5200 mods and upgrades

Then; discovered this thread.
Would any other CPUs work on this interposer for PGA132 386SX/DX socket interposer-based upgrades; such as that for the T5200?
Such as any of these below as "interposer candidates for enthusiast-built 386SX/DX PGA132 upgrade CPUs?"

1) AMD 5x86 (P75)
Full Model Number: AMD Am5x86-P75ADW
Clock Speed: 100 MHz (4x multiplier on a 25 MHz bus)
Cache: 16KB L1 cache
Bus Speed: 25MHz bus; but works with all bus speeds under 33MHz

2. Cyrix 5x86 (Cx5x86)
Full Model Number: Cyrix Cx5x86-GP100
Clock Speed: 100 MHz
Cache: 16KB L1 cache
Bus Speed: Optimized for 25 MHz or 33 MHz bus speeds but should work in the 20MHz bus of the T5200 via interposer.

3. Intel Pentium OverDrive

Full Model Number(s): PODP5V83; PODP3V83
Clock Speed: 83 MHz
Cache: 16KB L1 cache
Bus Speed: Supports 25 MHz or 33 MHz system buses. Internally, the CPU runs at a 2.5x multiplier when used on a 33 MHz bus.
Note: May not work on 20MHz bus without further investigation and modification

4. Cyrix 6x86 (M1)
Full Model Number: Cyrix 6x86MX PR166+
Clock Speed: 133 MHz
Cache: 16KB L1 cache
Bus Speed: Typically used with a 33 MHz bus, but can also work on lower-speed buses with clock multipliers.
Key Feature: The 6x86 series was designed to outperform Intel’s Pentium processors in certain integer-heavy workloads, with a robust cache to manage performance even on older system buses.

5. IBM 6x86L (Low Power)
Full Model Number: IBM 6x86L PR150+
Clock Speed: 120 MHz
Cache: 16KB L1 cache
Bus Speed: Optimized for a 33 MHz bus, but it’s compatible with slower bus speeds using internal multipliers (like the T5200)

I'm still building an interposer that *may* run those first three. The other two are Pentium class.

The interposer has a clock divider, disconnectable lines for suspend/FPU (SUSP, SUSPA, SMADS, PEREQ, BUSY, ERR), NMI select (between SX, DX).

I've built the new version in ENIG, but the press-fit 168 pin socket is a one-time only fit, so I need to order a new one before I can test it.

Small update:

I have been using this plug-in clock generator for 386 systems, shown here: Re: Project: Full Can Clock Oscillator Replacement It contains a programmable SI5351 PLL chip and I can set it to any frequency, up to 200 MHz. I believe the precision goes to at least the thousandths place, e.g. 85.000 MHz. This allows me to fill the void between 80 and 100 MHz since DIP-14 crystal oscillator packages don't normally exist in this range.

I spent some time go through my TI 486SXL2-66 CPUs to see where they stand.

Sample A
80 MHz: 3.85 V
87 MHz: not tested
88.5 MHz: 5.05 V
90 MHz: 5.20 V (wasn't quite stable during summer months inside a computer case)

Sample B
80 MHz: 3.95 V
85 MHz: 4.45 V
87 MHz: 4.75 V
88 MHz: 4.90 V

Sample C:
80 MHz: 4.0 V
85 MHz: 4.55 V
87 MHz: not tested
88 MHz: 4.90 V

Sample D:
80 MHz: 4.0 V
85 MHz: 4.55 V
87 MHz: 4.80 V
88 MHz: 5.0 V

Sample E:
80 MHz: 4.05 V
82 MHz: 4.25 V
85 MHz: 4.65 V
87 MHz: 4.90 V
88 MHz: 5.0 V

Sample F:
80 MHz: 4.05 V
85 MHz: 4.65 V
87 MHz: 4.90 V
88 MHz: 5.0 V

Sample G:
80 MHz: 4.05 V
85 MHz: cannot

Sample H:
80 MHz: 4.10 V
85 MHz: 4.70 V
87 MHz: 5.0 V

Sample I:
80 MHz: 4.10 V
85 MHz: 4.70 V
87 MHz: cannot

Sample J:
80 MHz: 4.20 V
85 MHz: 4.75 V
87 MHz: cannot

Sample K:
75 MHz: 3.65 V
80 MHz: Hangs at any voltage

Sample L:
75 MHz: 4.15 V
80 MHz: Hangs at any voltage

BEGIN EDIT: The above specs were later updated for more voltages/frequencies. END EDIT

I was primarily focused on testing at 85 MHz. 75% of the samples tested (9 of 12) were able to achieve stable 85 MHz operation in Windows, with voltages ranging between 4.45 V and 4.75 V. I also tested a few samples for 87 MHz, but not all. 3 samples were able to 87 MHz, but I suspect 6 samples, or 50%, should be able to do 87 MHz. One special sample was able to do 88.5 MHz at 5.05 V. This is the CPU which was able to run stable at 90 MHz during winter months inside a computer case, but warmer summer testing revealed it wasn't entirely stable.

It was interesting to see how low voltage operation at 80 MHz didn't necessarily translate to operation at 85 MHz (Sample G). Also how a very low operating voltage at 75 MHz (3.65 V) didn't scale at all, meaning this particular CPU wasn't able to do 80 MHz at any voltage (Sample K).

Have you done any tests with faster RAM, to see if there's a read/write speed issue?

I'm testing a 16-bit 486SXLC2 (386sx) board and 85Mhz/42.5MHz is exactly the region where a CAS Extend wait state is needed (write wait). 85MHz or below doesn't need it.

Do frequencies over 85Mhz, but with an extra DRAM wait state add stability?

My DRAM write wait options on this particular motherboard are 0ws, 1ws, and 2ws. If the issue was memory timings, why would 3 or more CPUs work fine at 87 MHz? I could test more if you wish, but for the case of 90 MHz during warm summer months, I recall trying to set 2 ws, but it didn't help the situation.

What clock generation unit are you using which allows for 85 MHz?

The thought was 87MHz is only slightly above, and the 88.5-90MHz tests may output more than 5V on the data lines which might make the RAM more stable.

There are actual 85Mhz and 90Mhz crystals on aliexpress. Nothing inbetween though that I saw.

If you look earlier in the thread I have a AnyClk device that allows apprx 1mhz increments.