feipoa wrote on 2025-11-04, 12:54:

The issue with Cyrix 5x86-120 chips and the 1024K module have been resolved. It seems that IBM 5x86c-100 chips soldered onto Th […]
Show full quote

The issue with Cyrix 5x86-120 chips and the 1024K module have been resolved. It seems that IBM 5x86c-100 chips soldered onto Thinkpad interposers are not well liked by the M919. Switching to some other interposers resolved the issues at 120 MHz with 1024K cache. This is odd because every other socket 3 motherboard I tested worked well with the Thinkpad interposers.

I was able to achieve a stable M919 with 1024K and Cx5x86-120 MHz with 2-1-2 and 0/0 ws. I used 8 ns SRAM (TSOP) and 50 ns EDO (TSOP).

Note that if you need to use 1/0 ws with EDO, you can get slightly better performance if you have some FPM that runs with 1/0 ws. For example, on the M919, and (I think) MB-8433UUD when running an Am5x86-180, Cachechk v7 as follows:

Copy code to clipboard

1EDO 0/0 ws --- memory read speed = 66.8 mb/s
2FPM 1/0 ws --- memory read speed = 55.4 mb/s
3EDO 1/0 ws --- memory read speed = 52.6 mb/s

And when running a Cx5x86-120, Cachechk v7 as follows:

Copy code to clipboard

1EDO 0/0 ws --- memory read speed = 55.2 mb/s
2FPM 0/0 ws --- memory read speed = 55.2 mb/s
3FPM 1/0 ws --- memory read speed = 43.9 mb/s
4EDO 1/0 ws --- memory read speed = 41.3 mb/s

Very, very impressive, I didn't think anyone would get one to run at 2-1-2 at 60mhz. Once again, you did it 😎

Was this with one of the latest s1R3 qfp CPUs or one of the older ones you had, or does the revision seem to affect the capability to run the tight timings at all? Did you also try with a regular IBM pga chip?

The 8ns sram must be giving you an extra edge here, especially in the tag

Is it windows stable is the next question 😀

I was testing the Cyrix at 3x40, not 2x60. Is 2x60 what you are after on the M919? I doubt I will get 2-1-2 at 2x60 on the Cyrix, but will try. Cyrix at 3x40, 2-1-2, 0/0 ws, 1024K was stable in Win95c once I used the Chinese interposers.

8ns L2 and 50 ns TSOPs are the best combination so far. I'm using 64 MB modules. With tight timings, cannot go to 128 MB even if it is 50 ns and TSOP. 60 ns TSOP Samsung doesn't cut it for the fringe cases.

I only have S0R5 on the Thinkpad interposers, so I cannot say whether it was S0R5 in connection with the Thinkpad interposers that the M919 doesn't favour. I have, both, S0R5 and S1R3 on the Chinese generic interposers and they run fine on the M919.

I played around with my four S1R3 QFP CPUs. I'd probably start a new thread if I get some success. So far, one of the CPUs stands out at 150 MHz on the M919, but I am getting the feeling that the M919 isn't the best board for these tests. I will switch to the UUD, M918, or LSD shortly. The S1R3 chips don't like high voltages. At 150 Mhz, you cannot go above 3.8 V, and cannot really go below 3.65V. There's a really narrow sweet spot, at least on the M919. All 4 chips turn on at 150 MHz, and I suspect, in the least, they can do 2x66 (hopefully with all features enabled).

EDIT: I can confirm that a Cx5x86-120 at 2x60 with EDO needs 3-2-3, 1/0 to be fully stable. FPM also works with 3-2-3, 1/0 and has a slightly faster read speed of 62.1 MB/s vs. 58.6 MB/s for EDO.

Anyone have an extra 1MB cache stick available for purchase? My post count isn't high enough to send a PM. There's some 256KB sticks on ebay, but I know I'd never be satisfied with that.

What's the market value on these M919 1024K sticks?

feipoa wrote on 2026-03-11, 13:06:

What's the market value on these M919 1024K sticks?

As usual in this hobby - whatever someone is willing to pay, especially since globally there is probably only a few 10s of them in circulation.

Around 2 years ago I think I spent around $120 CAD on one (that included shipping). I felt this was a tad on the high side, however since there was zero availability otherwise - I jumped on it. It wasn't even technically for sale - an ebay seller had an m919 board with a 1mb module inserted in it listed (for a lot more then I could afford). Upon asking them if they would sell the cache separately they replied that they wouldn't, however, they said they had a friend somewhere that "may have another module that they ended up not using available" and said they would ask this person to mail it to them if I'm interested. Sure enough, a month later they contacted me about the module now being in their posession and available to purchase.

With the rarity in mind I'd consider what I paid fair. We aren't really talking vintage pc parts here anymore, but rather artisan small scale production. The buyer isn't paying for something that has 30+ years of depreciation, but for something that someone personally developed, toiled upon, tested etc.

cluster.fsck wrote on 2026-03-09, 17:11:

Anyone have an extra 1MB cache stick available for purchase? My post count isn't high enough to send a PM. There's some 256KB sticks on ebay, but I know I'd never be satisfied with that.

If you have no cache right now I would recommend one of the currently available 256k sticks anyway. Performance will definitely be better than without. The market is poor right now and the price on those is good compared to what I have seen them sell for historically - if you eventually track down a 1mb one you should have no problem selling the 256k for what you paid.

With the time and effort required to acquire parts, assemble, and test, I'd say $120 CAD ($88 USD) is a bargain. The PCB's I ordered with 2U gold contacts from JLCPCB weren't cheap. The SRAM needed for these isn't sold through regular channels (digikey/mouser) any longer, so you often end up with junk that doesn't work and must be shuffled out and replaced. Extended testing is then needed to ensure the RAM holds its own. Then there's assembly time. The better SRAM for these is TSOP, but the leads are a bit long for the pads. As such, manual (slow) soldering seemed to work best.

feipoa wrote on 2026-03-11, 21:00:

With the time and effort required to acquire parts, assemble, and test, I'd say $120 CAD ($88 USD) is a bargain. The PCB's I ordered with 2U gold contacts from JLCPCB weren't cheap. The SRAM needed for these isn't sold through regular channels (digikey/mouser) any longer

Will these not work? https://www.digikey.com/en/products/detail/in … -10VXIT/1839251 I'm going off the list on
pancakepuppy's GitHub. $4.98 each (might as well get 10), so not cheap but won't have to worry about counterfeit chips.

Do you have any bare PCBs left? I'd be interested in one, provided the chips linked above will work. I was planning on going the JLCPCB route myself if I came up empty here.

I tried CY7C1019DV33-10 chips, both in SOJ and TSOP, but neither worked. Re: Lets make new M919 Cache sticks?

I recall one other forum poster tried Cypress chips, but also couldn't get them working.

For the $65 digikey cost, you can give it a 3rd attempt and order the IC's you linked.

If you are ordering from JLCPCB, rumour has it that their gold is too thin, so you might want to order ENIG with 2U gold instead of 1U.

Bare PCB's? I don't recall, but I don't think so. Expect around $60 shipped for 5 ENIG 2U PCB's from JLCPCB.

At some point, I'll be giving away my extra modules in exchange for registered charitable donations. Unfortunately, I'm busy with other things now and wouldn't be shipping to USA anyway.

feipoa wrote on 2026-03-11, 21:00:

With the time and effort required to acquire parts, assemble, and test, I'd say $120 CAD ($88 USD) is a bargain. The PCB's I ordered with 2U gold contacts from JLCPCB weren't cheap. The SRAM needed for these isn't sold through regular channels (digikey/mouser) any longer, so you often end up with junk that doesn't work and must be shuffled out and replaced. Extended testing is then needed to ensure the RAM holds its own. Then there's assembly time. The better SRAM for these is TSOP, but the leads are a bit long for the pads. As such, manual (slow) soldering seemed to work best.

I'm considering one of these contraptions on my m918 to try and find an 8ns SOJ tag that could run 2-1-1-1 with the 150mhz cyrix. Sure, it's actually not meant for SOJ but looking at some dimensions/how the pins connect it looks like it would work.

Could something like this work for finding cream of the crop sram without soldering? Just pop it in and test it as a tag (in mobos with dip sockets obviously)? Use sop32-dip32 variant for the higher capacity sram... I suppose the real question is if the longer leads to the IC would make it inherently unsuitable for this application, or, if this is not much of an issue at sub 66mhz fsb frequency

Assuming the surface mount pins will line up in the socket, I suspect you won't get very good results with that out-of-plane unit. But, the M918 is a fickle board, so maybe it will work just fine. In the least, you should be able to single out which SOJ IC's are best from a batch of chips that you test. Which part numbers were you looking for? Keep in mind that there are 300 mil and 400 mil SOJ's.

Your photo is shows a DIP 28 to SOJ 28. Wouldn't you need DIP 32 to SOJ 32 for 1024K SRAM? Here's a 64kx8 8 ns SOJ 300 mil I have. I think the 'L' is for low power though, meaning 3.3 V.
download/file.php?id=165930&mode=view

Keep in mind that inserting those squared pins into an SRAM DIP socket will destroy the DIP socket, assuming you can even force it in. Either you'd want to replace the pins on that adaptor with thinner pins, or be prepared to replace the DIP socket on your M918.

feipoa wrote on 2026-03-17, 02:16:

Assuming the surface mount pins will line up in the socket, I suspect you won't get very good results with that out-of-plane uni […]
Show full quote

Assuming the surface mount pins will line up in the socket, I suspect you won't get very good results with that out-of-plane unit. But, the M918 is a fickle board, so maybe it will work just fine. In the least, you should be able to single out which SOJ IC's are best from a batch of chips that you test. Which part numbers were you looking for? Keep in mind that there are 300 mil and 400 mil SOJ's.

Your photo is shows a DIP 28 to SOJ 28. Wouldn't you need DIP 32 to SOJ 32 for 1024K SRAM? Here's a 64kx8 8 ns SOJ 300 mil I have. I think the 'L' is for low power though, meaning 3.3 V.
download/file.php?id=165930&mode=view

Keep in mind that inserting those squared pins into an SRAM DIP socket will destroy the DIP socket, assuming you can even force it in. Either you'd want to replace the pins on that adaptor with thinner pins, or be prepared to replace the DIP socket on your M918.

I was actually looking at 32kx8 8ns of those Elites you posted

Right now I'm just after a tag, since on this board I'm using 512kb of cache- I have accumulated a few sets of original srams in that capacity in dip32. They are of course all 15ns, but also all perform better then anything chinese/relabeled I have.

BUT! I think the main factor here is the tag speed, likely even more so with the cyrix. I have several Chinese looking ISSI 32kx8 -12, and now some genuine looking Winbond 32kx8 also -12. All these work fine with any combination of of other sram I have at 3-1-1-1 at 50mhz, or even 60mhz for that matter...But ONLY ONE of the ISSIs will make it as far as dosbench at 2-1-1-1 before failing in tests there - so I'm thinking it's the better chip the bunch and I should probably be looking at the fastest speed grade that I can get

Edit: right, the pins - you are right, all these adapters have fat pins - I'll use an unsoldered socket with thin pins in between and ram the adapter into it instead so I don't wreck the mobo. I ordered 1 cheap (12$ cad) solder on soic28 to dip28 adapter from digikey that should work ok in my application - it's a little wider than dip 28 but with the extra socket in between it won't matter.