I updated the interposer I'm working on to better suit all 486 CPUs, and updated the PGA-168 press-fit holes on the PCB. Currently saving up for another prototype.

Having long-suffering problems with the press-fit socket and the PCB finish. I think I have it worked out though and need at least one more prototype run.

List of changes:
LT1573 voltage regulator & MJD45H11 PNP
Removed NA# jumper and added NA# solder-jumper to reverse side.
Added FLUSH# logic (MEMW AND'd with inverted HLDA) to suit both TI486SXL and 486SX/DX.
Added solder-jumpers for R17 & B13 to GND (for CLKMUL#). 20k pullup to 5v for R17.
Added 20k pullup to 5v for FLT#/UP#/RESERVED# (C11). Specified for intel CPUs.
Added 20k pullup to 5v for DP0-DP3 (N3, F1, H3, A5).
Smoothed CLK/CLK2 lines and reduced travel distance
Reduced PGA-168 holes from 1.15 to 1.0
Cosmetic changes

MSxyz wrote on 2025-05-17, 07:55:

The short answer is that I'm looking for an interposer that will allow me to use the PGA-168 version of the 486SXL2 (like those sold on eBay) on any 386 motherboard that already supports a Cyrix 486DLC (I've a few of them, all tested working).

Even if your motherboard BIOS supports the DLC's cache, you will still want to use cyrix.exe to make register adjustments, depending on your particular hardware, and if you are running Windows. For the most part, the existing alpha1/2 interposers should work with most traditional 386 motherboards. It has worked in about a dozen motherboards I have tested. There was one fellow who had some IBM type motherboard and had to add some additional cache flush logic, but these options are included in both PCB revisions if you add the Q1 and Q2 chips.

MikeSG wrote on 2025-05-17, 08:00:

I updated the interposer I'm working on to better suit all 486 CPUs, and updated the PGA-168 press-fit holes on the PCB. Currently saving up for another prototype.

Having long-suffering problems with the press-fit socket and the PCB finish. I think I have it worked out though and need at least one more prototype run.

If it is only one more prototype run, that would be impressive. The wish list for your interposer is much more involved than a simple SXL132 to SXL168 interposer. I am looking forward to testing some fast AMD 486SX2 processors against the fastest achievable SXL2 (currently 90 MHz).

For anyone else wanting a pre-assembled unit, please wait until I have a batch of these assembled (probably November). One member made such a large donation [to a charity I selected], I felt more compelled to assemble a one-off. The problem with this is the doubling of time required. For the one-off, this is what my tracked time looked like this:

1.5 hrs: prep, e.g. pulling parts and cutting & sanding down pin strips
7.5 hrs: soldering socket, pins, and filtering caps
2 hrs: all other soldering + flux cleanup
1 hr: setup a testbed, run scope measurements, take photos, and disassemble the testbed.

Total: 12 hours.

I hope to get this down to 4-6 hours each unit in a batch, but its hard to say. Soldering the pins and socket takes the most time. The amount of exposed conductor on the PCB is quite small and makes heating more difficult. I needed to use a finer tip, which is OK until you reach a GND or VCC pin. Bent tips for the corners are best.

Unfortunately, the guy selling the narrow PGA168 sockets is now asking about $6 each (shipped). They used to be about $1.50.

I reduced the number of caps to solder on, although feel the quantity is still excessive. This was my prettiest assembly since I removed the no-clean flux, didn't have caps to swap out (over and over again), and didn't have burn marks. I soldered the bottom layer 0805 caps down at the pin base to look nicer at the expensive of more time. Some photos:

Looking nice

I think it only took 1-2 hours to solder my last prototype. But yours is working and mine isnt...

About the caps. I'm swapping the voltage regulator on my design over to a LT1573 (with a MJD45H11 PNP) and in the manual there's a specific setup for 3.3v 5Amp microprocessors. It recommends 100uf on each side of the regulator (5v side, 3.3v side), and ~24 1uf caps.

So all of the 0.1uf caps on the CPUs VCC-GND I switched over to 1uf, with 33uf on each side of the voltage regulator. I don't think they need 0.1uf on each VCC-GND pin.

Nice! :p

Yeah, it took me forever to get mine assembled. It was quite the build. So tiny….such tight spaces… Lol.

Indeed test fitting and clearancing the lower pins is critical! 😀

Do I recall you saying that the older version gets better overclocking? Is that still the consensus?

It would be cool to try and figure out why and update the latest version to match or exceed that. It’s been so long since I made it I don’t recall all the details. But those reading, I always suggest the latest version for building. 😀

Thanks to Feipoa's kindness and patience, I was finally able to run one of those 486SXL2 in my collection. 😁

I went straight to what I thought it was the fastest motherboard in my collection that could make use of this CPU, the (Daewoo?) Panda 386V. It's based on the ALi 1429/31 chipset and it's one of those rare 386 motherboards with a true VLB slot. But it turns out I was not entirely correct. While this board is as fast as they come, I haven't been able to make the 486SXL2 work nicely with a FPU - any FPU in clock doubled mode. The BIOS (dated 22 feb 1994) recognizes the Texas chip and the board also supports Cyrix 486DLC including the internal cache and the necessary control signals. For some reason, however, once I enable clock doubling via the CPU status register, the system hangs as soon as I launch some computation intensive software.

I've tried messing around with different FPUs, lowering the clock frequency and raising wait states and changing other 'bits' in the CPU status register. Nothing really worked! Without the FPU, however, the system appears to be rock solid even at 80Mhz. It completed my DOS port of the C-Ray32 benchmark in 3h42' several times in a row without hanging (the lack of a FPU really shows here...). And with some mild cooling (a 12cm fan blowing air on the motherboard) the CPU stays barely warm to the touch despite a necessary overvolt to 4.0V to make it stable at 80MHz

The performance is really impressive for a 386 derivative... With the Doom 1.9 timedemo playing at full screen minus the status bar, I get 3138 realticks... That's 23.8FPS! It's almost as fast as a 486DX-50; memory and throughput performance are equally impressive. A shame about the FPU, really... I've now moved the chip to another motherboard (Soyo 019R1) which is based on the same ALi chipset and it's seems the FPU issues are gone... Yesterday evening I played the Quake demos in loop for a couple of hours without problems. This board, however, has only ISA slots..

Nice, you have it working! I sent you a hacked BIOS for that Panda 386V board; it will hopefully resolve your FPU issue.

On your Soyo board, were you able to determine the minimum CPU voltage to yield 80 MHz operation?

I don't have a VLB SXL2 system setup, but it would make for a nice addition. Holding me back are the PLL limitations on these Panda boards - they top out at 80 MHz. It is on my list is to bypass the PLL to yield something closer to 85-88 MHz.

feipoa wrote on 2025-06-14, 04:23:

On your Soyo board, were you able to determine the minimum CPU voltage to yield 80 MHz operation?

I haven't touched the potentiometer since I moved the CPU to the new board. At 4.0V the system is stable and able to run the usual tests in loop for hours. I haven't tested the other 486SXL2 in my collection yet. Since I've found a stable configuration and put all my problems with the FPU past me, I'll keep this system running in this configuration for a few days more.

The lack of a VLB card is felt, though... These are the DOOM 1.9 timedemo scores at various ISA clocks (as always in full screen minus the status bar at the bottom):

ISA Clock Realticks FPS
8 MHz 4719 15.8
10 MHz 4315 17.3
13.3 MHz 4105 18.2
16 MHz 3826 19.5
20 MHz 3657 20.4

Not even at 20 MHz I'm able to approach the score of the other board (23.8 FPS) despite obtaining similar performance in benchmarks where graphics aren't a factor. In a RL scenario, 20 MHz would also a bit too high to allow stable operation of other cards such as network or sound. From my experience 16 Mhz would be the maximum stable with most boards; older ones might even start to "panic" at speed above 13 MHz 😀

pshipkov is running his 386/486 system with a 33 Mhz ISA bus. Said sound and network works fine. I know, it's hard to believe:
Re: 3 (+3 more) retro battle stations

I haven't had his luck, but apparently, the luck is in the pickings of the ISA cards.

feipoa wrote on 2025-06-14, 07:38:

pshipkov is running his 386/486 system with a 33 Mhz ISA bus. Said sound and network works fine. I know, it's hard to believe:
Re: 3 (+3 more) retro battle stations

I haven't had his luck, but apparently, the luck is in the pickings of the ISA cards.

Out of all cards, the VGAs seem to have the least problem with a turbocharged ISA. The Tseng ET4000AX, despite being a 1990 chip was used also in some early VLB cards and I never had a problem running those chips on a 20 MHz ISA bus. I own some late ISA cards based on the CL-GD5434, Tseng ET4000W32, WD90C30 and these also work reliably at 20 MHz.

As for Multi I/O, I find that chips made in 1993-94, and which were also used in VLB boards , are also a safe choice. Currently I use a pair of Acer SAB-560 Maximum AT boards and these seems to be tolerant of high speeds as well. Sound cards are more picky. Recently, I've re-discovered the C-Media sound chips and I'm currently using a board from an unknown manufacturer with the CM8330; this also seems to be tolerant of 20 MHz operation. I've no experience with network or SCSI cards.

Unfortunately, the Soyo 019R1 won't let me run the ISA bus at more than 20 MHz, to test its stability at higher speeds. The BIOS uses the system clock as source and has a wide range of dividers (in addition to the "safe" choice of 7.15MHz), but it offers no lower ratio than /4, so I can go as high as 20 MHz at 80MHz and 16.67MHz at 66MHz.

In the previously linked thread, pshipkov is using VLB for graphics. IDE is through an integrated VLB controller. Nonetheless, on pure ISA boards, he also found that when running ISA at 25 MHz, with the right motherboard and controller cards, you could often get better than VLB results for the same CPU.

Some motherboards are well optimised for the SXL2 CPU, while other boards lend preference to the IBM BL3.

For my systems, I usually add some components which tend to limit my ISA speeds, like SCSI, an LS-120 ISA card, 3Com Etherlink III, or other novelties. Because of this, you won't normally see top-end results from me. For example, on my SXL2 system, running at 88.5 MHz, I could not achieve stable ISA speeds over 11.06 MHz. This yields a DOOM result of only 3434 realtics, or 21.75 MHz. I think 24 fps is achievable with faster ISA speeds, especially on a GD-5434.

Did you measure your ISA CLK signal with an oscilliscope? Sometimes BIOSes take the crystal oscillator value and divide by the displayed BIOS divisor value, and sometimes they take crystal/2, then divide by the displayed BIOS divisor. The naming convention they choose is not always correct or standardised and I find the best approach is to measure the ISA CLK with an oscilliscope. Then you can make a note in your manual to clarify if they are using CLK or CLK2 as their convention.

No I didn't measure the bus with the oscilloscope, but I've tested this board against others before and also compared the results when picking the 'safe' 7.15MHz option and I'm quite sure it uses the system clock.