Did you read above? I was thinking about 10, 100, 200 nF here. But no feedback. Tantalums of size 2312 do not come in these capacitance values. Smallest tantalum in 2312 is 470 nF.

We might be able to use 1210 ceramics (3.3x2.5 mm) but the centre section is around 2.1 mm and the PCB gap is 2.85 mm. I can hover the caps a bit to not touch the centre gap, but the overhang of the metal contact would only be 225 um thick on each side.

This is what I'll be ordering for caps. I think it should cover any changes I might make after testing and any recommendations which may come in from viewers later on.

Tantalum 2312 6.0 mm x 3.2 mm
0.47 uF
10 uF
22 uF
47 uF
100 uF

Tantalum 1411 3.50 mm x 2.80 mm
150 nF
220 nF
330 nF

Ceramic 1210 3.30 x 2.50 mm
10 nF
47 nF
100 nF
220 nF

Price of caps have really gone up.

I think at least one of the center cap pads has nothing between vcc and gnd, so should be no problem slapping small smd ones there even if Sphere478 doesnt extend them.

Use kapton tape.

If you guys haven’t used that stuff before you need to check it out. Tape that doesn’t melt when solder is on it or iron. Great stuff.

feipoa wrote on 2022-10-06, 05:59:

I think I'll look for some thru-hole caps in a package similar to that of conventional resistors.

Wouldn't use through hole for the nF range capacitors. The pins will have too much inductance, effectively canceling out the benefits of the small, high frequency capacities..

Not sure if this is useful to anybody here or not.
https://github.com/wiretap-retro/486-Pentium- … ocket-for-KiCAD

Kapton tape - do you have a link for the item you use?

https://www.google.com/search?rlz=1CDGOYI_enU … 5&bih=640&dpr=3

Just a google shopping search^

It comes in all sizes. Just buy whatever size you like and cut a piece that fits.

just fyi kapton is a brand name, but nowadays you can buy Chinese kapton and almost kapton kapton 😀

If anyone has questions about how to order. Here is a quote from a conversation in PM with some added commentary:

The gerber zip in the zip file is what you upload.

The aluminum option isn’t the one you want. Just use default pcb type. I think aluminum only works as one layer anyway.

Jlcpcb is good about checking it before making it. I say order it as it uploads it. Might save money.

Talk to them in the online chat feature if you want they are pretty helpful usually.

Yeah I mentioned that it only detected 4 layers. Jlc’s online pcb preview software is glitchy. But the product always seems to come in correct in physical form.

Pcb thickness can be whatever you want. 6 layer will only have a few options though. So on another tab see what it lets you do when 6 layer is selected. I’d say for this project go 1.6 unless you think that would be too thick for some reason

Confirm file yes might be helpful wouldn’t hurt…

Kelvin wire default

Gold finger no

Material whatever default is

Copper weight, whatever default is. (We don’t need to go extra thick)

Gold thickness default (only gold if you select gold finger and the more expensive production method. There is no need for this)

Castellated holes. No

Appearance quality default.

Hole diameter whatever default is. I didn’t use any minimum size holes so it won’t matter.

Silk default

Layer stackup default

Delivery format default

Different design default

Impedance default

Product type default.

This is a prototype. Just choose the default cheap options for most of this. Really the only thing you should mess with is pcb thickness and color.

If you want super high quality order that for version pi.0 the final version. But you’ll probably decide that the cheap options are more than adequate for what we are doing.

It won't matter if I leave the layers as 4 layer, or move it to 6 layer? If I select 4-layer, the cost is $28. If I select 6 layer, cost is $86. If the product comes out correct either way, wouldn't we always select 4-layer?

PC thickness: If I select 2.0 mm instead of 1.6 mm, might this increase the distance between the inner layer traces? If so, this seems more ideal with respect to possible cross-talk. The machine pin leads are 3 mm, so 2 mm wouldn't be too thick.

But there's a good sized price difference for 2 mm vs. 1.6 mm, that is $139 for 10 at 2mm vs. $86 for 10 at 1.6mm

thicker also means further from the ground plane. The closer a signal trace is to its reference ground the better, the further the more it radiates.
btw for your experiments with that other converter - instead of weird aluminum cans you could just glue a flat piece of alu foil over the pcb. Oobviously with insulation layer between, electrical tape will suffice.

feipoa wrote on 2022-10-08, 10:46:

It won't matter if I leave the layers as 4 layer, or move it to 6 layer? If I select 4-layer, the cost is $28. If I select 6 layer, cost is $86. If the product comes out correct either way, wouldn't we always select 4-layer?

PC thickness: If I select 2.0 mm instead of 1.6 mm, might this increase the distance between the inner layer traces? If so, this seems more ideal with respect to possible cross-talk. The machine pin leads are 3 mm, so 2 mm wouldn't be too thick.

But there's a good sized price difference for 2 mm vs. 1.6 mm, that is $139 for 10 at 2mm vs. $86 for 10 at 1.6mm

That’s what I’m saying. Send it through as detected 🤣.

They may fix it later and not charge extra.🤷‍♂️

Most likely worst case is they email you asking for more. But they are pretty good about emailing you about problems. I’d say chances are that it comes in correct. Your call though. If you want to go by book select 6 layer.

They don’t do many 2.0mm so it is more expensive. But your logic is probably sound with spacing assumptions.

It’s a prototype. Order 5 at 1.6, choose the cheapest color. Your call on the layers thing. Might be able to cheat the system, may come back to bite you. If you are conserned. Ask them on chat.

OK.

Is the option "Confirm Production file: Yes, No" asking a human at JLCPCB to look at the gerbers for errors, then flag the customer before production? If so, seems like we'd not want to check this.

Yes, you can do that. Go for it 😀

I've got a small update from my side-by-side interposer. As I didn't get the cache working, I wanted to hook up MEMW in order to be able to use cache flush. Normally, this is one of the pins of the ISA bus, however I have an IBM with only microchannel slots, and MCA doesn't have a MEMW pin.
After digging through the MCA specifications I found that two pins are used to the same effect: S0 (pin A32 on the MCA bus) is WRITE, and MIO (pin A34) toggles between memory and I/O. The combination of this would be equivalent to MEMW.

MEMW is active low, so I needed a circuit that did: MIO high + S0 low = low
For this I used a NAND inverter for MIO and feed the inverted signal and S0 into an OR gate, and hook the output to MEMW.

I made a little pcb and tacked it to the interposer:

With [CYRIX.EXE -f -m -cd -xA000,64] the system now reports a working cache:

The cache works in practice, I went from 9700 RTs in doom with the original 386DX-20 to 3100 RTs with this contraption, a three fold improvement. Unfortunately it is not rock solid; doom works fine, but command&conquer locks the computer during startup - UNLESS i run doom first, and then c&c :'-D and then it still crashes after half an hour.
I tried everything I could imagine (A20, first 64k, BARB, excluding different zones) but so far nothing had any effect on this behavior. I find it very mystifying; does doom leave something behind in memory?

Nice bit of detective work for finding a MEMW alternative on MCA.

That DOOM score of 3100 realtics is too good for an SXL2 at 40 MHz. Are you running it in full screen?

By the way, the new PCB's are on their way from Hong Kong, as of last Friday.

feipoa wrote on 2022-10-17, 11:40:

Nice bit of detective work for finding a MEMW alternative on MCA.

That DOOM score of 3100 realtics is too good for an SXL2 at 40 MHz. Are you running it in full screen?

By the way, the new PCB's are on their way from Hong Kong, as of last Friday.

I'm running the low detail benchmark haha 😀 I guess if I want to really play doom, I'll go with fdoom.
Super curious how the PCBs will turn out, please keep us posted!

Oh wow, very nice :p

Hope you figure out why there are instabilities. Maybe a non related problem is rearing its head? Or maybe a timing issue? Are there faster components you can try?

Btw, I may have enough room for that on the back side of the interposer.

Can you upload the kicad file? I can try to copy it to the interposer.

ram? both doom and c&c are dos4gw games, maybe running dos4gw twice in a row makes it allocate/load itself into different ram address?
try different dos4gw game

Sphere478 wrote on 2022-10-17, 15:04:

Oh wow, very nice :p

Hope you figure out why there are instabilities. Maybe a non related problem is rearing its head? Or maybe a timing issue? Are there faster components you can try?

I think its not related to the interposer. I had an 486DLC in this sytem and it showed the same instability, although it didn't need the MEMW to work. That processor can run its cache direct mapped, whereas the clock doubling ones are fixed on 2way associative. Unfortunately there is not much to tinker here, the IBM bios does not have any options for memory management, or easy speed changes.

Btw, I may have enough room for that on the back side of the interposer.

Can you upload the kicad file? I can try to copy it to the interposer.

That's pretty cool, but do keep in mind almost nobody is going to use this option, so make sure its optional and people do not need to components to use a normal MEMW signal.
In any care, here you go: