pshipkov wrote on 2025-01-12, 19:54:

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

It's a TI486SXLC2-G50. I believe they are 3.3V-5V compatible. So very stable at the standard 386sx 5V. But do require a heatsink & fan. And can't be soldered with a hot air gun.

It's a lot less transistors than the 32-bit version so that may help with getting to 100Mhz. If there are XTALs above 100Mhz and the keyboard worked I'd try higher...

MikeSG wrote on 2025-01-13, 08:15:

It's a TI486SXLC2-G50. I believe they are 3.3V-5V compatible. So very stable at the standard 386sx 5V. But do require a heatsink & fan. And can't be soldered with a hot air gun.

It's a lot less transistors than the 32-bit version so that may help with getting to 100Mhz. If there are XTALs above 100Mhz and the keyboard worked I'd try higher...

That's good to know the SXLC2-G50 can push 100 MHz, and possibly higher. I guess we don't need to hold out for the rare SXLC2-66 variant. How many keyboard controllers did you try? Why is there a limitation on hot air?

feipoa wrote on 2025-01-11, 10:48:

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.

To answer my own question, I ran a few tests with the switching regs on these $1 buck boards (MP14820S) and the dropout is around 2V at load. If Vin to the Buck is 5.0 V, the highest Vout [from the Buck] I could get was 3.03 V. This was with a load of around 600-800 mA. Thus, it looks like we cannot merely use these buck boards in place of the linear LDO, unless they are wired to the 12 V rail.

Yes, OUT to OUT. Fixed in my previous reply.

No SXLC2-66 here. Only 50MHz rated ones.

It will take time to determine if the different hardware configurations make any difference.
The current setup vs buck only and the interposer stripped vs presence or absence of caps on the interposer.
Not sure i am willing to spend the time for that. Not sure it will make a difference really.

feipoa wrote on 2025-01-13, 08:31:

That's good to know the SXLC2-G50 can push 100 MHz, and possibly higher. I guess we don't need to hold out for the rare SXLC2-66 variant. How many keyboard controllers did you try? Why is there a limitation on hot air?

Tried an Intel P8242-AH (phoenix) which was too slow. Then a VIA VT82C42. Thinking about testing a Winbond W83C42. The last two are fast/hardware implemented, but ... working through the difference in requirements. What the board needs... Changed the 74F04 to a 74F06. Just need to go through each line.

Everyone I see comment about their SXLC2, can't double clock the chip at all... But both chips I've soldered onto two different motherboards (hand-soldered) worked straight away no problems. 66Mhz minimum for both.

Re-ordered these gold PGA pins from another supplier. Luckily the product from the second supplier had the usual gold composition. Here's a comparison of the two. Something to watch out for if you are going to assemble these.

Nice, but I think that it can change everytime you order, sellers just re-sells from various manufacturers...
Where did you buy it from? I ordered 2 or 3x from Ali and had a luck I got decent gold coating...

I bought both packages from eBay.

Ideally, the seller who is selling the "thin gold" pins should complain to their supplier. If enough sellers are being forced to refund for pins with bogus gold, it will hopefully put pressure on the suppliers. For the seller with the bogus gold, he kept trying to assure me they are fine and have gold, even after I showed him the photos. eBay initiated a refund to me. I have no need for the gold-less pins, so will ultimately throw them out.

I ordered about 1200 pins, so I hopefully won't need to buy these any time soon.

The eBay listing for pins with bogus/thin gold: https://www.ebay.ca/itm/373634328180

The eBay listing for correct gold concentration: https://www.ebay.ca/itm/285532629187

MikeSG wrote on 2025-01-13, 18:09:

feipoa wrote on 2025-01-13, 08:31:

[...] How many keyboard controllers did you try? [...]

Tried an Intel P8242-AH (phoenix) which was too slow. Then a VIA VT82C42. Thinking about testing a Winbond W83C42. The last two are fast/hardware implemented, but ... working through the difference in requirements. What the board needs... Changed the 74F04 to a 74F06. Just need to go through each line.
[...]

About keyboard controllers at 50Mhz.

I tested a W83C42 keyboard controller today on the Ti486SXL(C)2, with a SN7407N open-collector type Hex Buffer IC. Works great.

Tested Duke Nukem 3D @ 100Mhz. 20Mhz ISA bus. Speeds up and slows down a lot. 25-30FPS if you face a wall. 10-15FPS normal. 3FPS if facing an explosion full screen.
Doom benchmark 21.8FPS.

Was stable. Haven't tested Win95 though.

Both 80MHz (no CAS delay) and 100MHz (CAS delay) give 20.8MB/s memory speed in Cachechk and seems to be the most stable.
85MHz gives 22.1MB/s but Duke3d crashes.

Memory speed seems to be 95-100% the bottleneck...
-3DBench only gives 5% higher scores when double clocking.
-Doom is the same whether it's double clocked or not...

So to continue I would need to test 45ns RAM modules with 90-100MHz.. for maybe 25MB/s. 2-3FPS.

A 386sx board with cache would be a lot better.

Huh, so the double clocking really doesn't do much of anything at that point. Interesting.

Have you generally seen sort of an asymptotic behavior with regard to the benefit of double clocking as you're increased the overclock?

It's just with Doom that I tested.

Doom:
40Mhz - 80Mhz CPU speed. Difference = 0.2FPS
42.5Mhz - 85Mhz CPU speed. Difference = 0.2FPS
45Mhz - 90Mhz CPU speed. Difference = 0.2FPS
50Mhz - 100Mhz CPU speed. Difference = 0.0FPS

3dbench is better (doesn't use RAM as much):
40Mhz - 80Mhz CPU speed. Difference = 2.7FPS
42.5Mhz - 85Mhz CPU speed. Difference = 2.4FPS
45Mhz - 90Mhz CPU speed. Difference = 2.4FPS
50Mhz - 100Mhz CPU speed. Difference = 1.8FPS

In Doom, at the same 20.8MB/s RAM speed. Changing the system clock from 40/80MHz to 50/100Mhz increases the Doom score from 19.2 to 21.8.

So I'd put the bottleneck as 1- Ram, 2 - System Clock. For the 386sx / Ti486 combo.

Interesting. On 386DX systems, the clock doubling helps quite a bit. Do you have more than one 386SX board to verify repeatability? I don't remember such poor results on my upgraded 286 board.

Speaking of cache... Just thinking aloud, but are there socket adapters with cache onboard for boards without it?

Think of it like a pentium2. The core "plugs in" (soldered) to the socket adapter (slot card) which also carries cache, and then plugs into the board.

They must exist... Right?
If not, I'm sure someone here has the brains to design one.
With more modern chips, full fsb speed and 1mb shouldn't be out of the question.

feipoa wrote on 2025-02-15, 19:36:

Do you have more than one 386SX board to verify repeatability?

Can't test on another board at the moment.

myne wrote on 2025-02-16, 05:42:

Speaking of cache... Just thinking aloud, but are there socket adapters with cache onboard for boards without it? […]
Show full quote

Speaking of cache... Just thinking aloud, but are there socket adapters with cache onboard for boards without it?

Think of it like a pentium2. The core "plugs in" (soldered) to the socket adapter (slot card) which also carries cache, and then plugs into the board.

They must exist... Right?
If not, I'm sure someone here has the brains to design one.
With more modern chips, full fsb speed and 1mb shouldn't be out of the question.

There was a 386sx socket produced at one point that sits over the top of the CPU, but... I don't think that socket is available anymore and don't know how the cache stays coherent.

If I was going to fabricate something, I'd make a whole new motherboard... The Chips F82C836 can be bought new, supports 50Mhz, and actually has write-back cache support.

I did test the Ti486SXL(c)2 with an ALI M1217 chipset some time ago. It maxed out memory speed (20-21 MB/s) at only 33/66Mhz. So at 40/80Mhz and with the EDO DRAM it supports, it can stay at minimum wait-states, and the FPU will work. But no cache support.

Cache makes 3dbench fly, but maybe only 5% difference in Doom/Duke3d...

I think you get:
3 FPS per 4Mhz ISA Bus Clock increase.
2 FPS per 10Mhz sys clock increase.
1 FPS per 1MB/s memory speed increase.
1 FPS for L2 cache (Doom/Duke).
0.2FPS for clock double (Doom/Duke).

I think Kingston made some 286 to 386SX upgrades which had some onboard SRAM. I haven't tested it though.

MikeSG wrote on 2025-02-17, 10:33:

If I was going to fabricate something, I'd make a whole new motherboard... The Chips F82C836 can be bought new, supports 50Mhz, and actually has write-back cache support.

Oh, is that that chip that they are putting in the new 386 laptops? it's a system on a chip 386 if I recall.

This makes me wonder if the 386 bus exits that chip, or if it's all internal. If it does, maybe there is a way to disable the onboard cpu and use an external one. like a sxl2-66. That would be cool to build new mobos with a good onboard cpu but a socket for using others.

a interposer that adds cache or memory would also be cool. (picoMem should enter the chat?)

Sphere478 wrote on 2025-02-18, 04:38:

Oh, is that that chip that they are putting in the new 386 laptops? it's a system on a chip 386 if I recall.

This makes me wonder if the 386 bus exits that chip, or if it's all internal. If it does, maybe there is a way to disable the onboard cpu and use an external one. like a sxl2-66. That would be cool to build new mobos with a good onboard cpu but a socket for using others.

a interposer that adds cache or memory would also be cool. (picoMem should enter the chat?)

I think that's the ALi M6117 (an M1217 + 386sx in one) - https://theretroweb.com/chipsets/182

Disabling the onboard CPU would just be an M1217 chip. It does support EDO DRAM, so you can have high memory speed w/o looking for 40-50ns FP DRAM...

Maybe somebody knowledgeable about cache can comment on whether you can just add cache w/o the chipset knowing. I know the Chips F82C836 has 3-4 BIOS options related to cache. Write-back, Early RDY return, Hidden Refresh...