pshipkov wrote on 2021-08-20, 18:23:

They for sure improved stability at 60 and 66MHz. Profiled this carefully.

Under which BIOS settings and which test conditions did 10 ns improve stability? I did not witness this, however, I did not extensively test it with 15 ns first. 3-1-3 and fastest worked fine with 15 ns for my initial rounds of tests.

when I solder SOJ I do it by hand

*pre tin all pads
*apply downward pressure on chip with finger while soldering
*lots of no clean flux

maxtherabbit wrote on 2021-08-20, 23:14:

when I solder SOJ I do it by hand […]
Show full quote

when I solder SOJ I do it by hand

*pre tin all pads
*apply downward pressure on chip with finger while soldering
*lots of no clean flux

"by hand", as in with hot air, or contact iron?

My approach for soldering:
1. Heat gun one element at a time, give the PCB time to heal (cool-down).
2. With 900C soldering iron chemwik gently (flash-clean) the pads - making sure they are 100% clean from any residuals.
3. Wipe the pads and surrounding surface with pincette and small ball of paper towel soaked in acetone. After that step the area should look like there was no soldering work done to it.
4. Run a line of soldering paste through each array of pads. Needs to be just the right amount. Otherwise it won't solder properly, or bulge in the wrong places.
5. Place the element properly. Press it with the back of pincette or other fine tool.
6. Start heat-gunning. Once the paste turns into tin after few seconds, gently press and hold the element down with pincette then move the heat-gun away. Keep like that for few seconds to form good contact with the pads underneat.
7. Repeat #3.

Sounds like a lot of steps, but it is actually very quick.

Here is what i got for the LS D upgrade:

---

LS D was not 100% stable at 3x60 with 2-1-2 L2 cache timings and default 15ns cache chips.
Worked well with the usual DOS/Windows tests.
Completely failed complex compute tests.
Solid with 10ns.

feipoa wrote on 2021-08-21, 00:17:

maxtherabbit wrote on 2021-08-20, 23:14:

when I solder SOJ I do it by hand […]
Show full quote

when I solder SOJ I do it by hand

*pre tin all pads
*apply downward pressure on chip with finger while soldering
*lots of no clean flux

"by hand", as in with hot air, or contact iron?

Contact

I see. So 10 ns SRAM allowed for stable operation at 60 MHz and 2-1-2? Was there any benefit at 66 MHz?

I always forget about solder wick. In my mind, removing all the solder from the existing pads would have been too time consuming, but in the end, not doing it required 10x more time than if I had removed it from the onset. I left the old solder on the pads, but smoothed them out with a quick run of the iron. However, they will never all be at the same height with this approach. Maybe 5-10 in 28 pins per IC will actually have the leads touching the tinned pads. On the remaining pins, I had to bridge that small air gap with solder, which is very difficult giving the angle and tight confines. Maybe there is a better method to get those tinned pads to a uniform height? The issue with SOJ is that the contact point is a bit under the IC, so you cannot really see the air gap well like you could with QFP, for example.

It was not fun, but turned out OK. Doing each pin one by one like this, that's 252 pins, and when I was finished, I guess I missed just one of these air gap bridges because it didn't work. I was able to hunt down the culprit with a multi-meter pretty quick though.

Here's what my mess looks like. I'll test my old solder paste before I butcher my NexGen board. I think using the desolder pump on the ZD-915 will be just about as fast as the wick, wouldn't it? I was thinking of using that to de-tin the pads.

With my technique above the whole procedure takes, i don't know - 20 minutes ?
Also, it takes care of the contacts point issue for good. Using the wick (assuming good one) is for sure better/faster than pump.
I actually tested after swapping every two chips. Turned out it was enough to swap the 2 upper right ones only (i actually started with them, since they are far away from plastic elements) in order to achieve complete stability + best cache wait states. So i could be done in 3 minutes, but decided to change the rest as well for better optics.

-

Yes, 10ns gets you to 2-1-2 at 60MHz and complete stability.
Didn't test enough 3x66 before moving to 10ns chips, so cannot comment if there was difference or not, but if 60MHz improved - i assume the same was true for 66 as well.

-

So, LuckyStar revision D with 10ns L2 cache transformed into the meanest motherboard at 180MHz.
Updated the original post with fresh numbers and notes. (may need hard reload of the page for the new images to show up - in case you visited them recently).
Until now i kind of expected that Biostar UUD will end-up in the PC case, but not anymore.
There are two more motherboards left on the list to examine - Chicony TK8880F (below) and PC-Chips M919 - will go over them for completeness, but they are just not in the same league and there is no room for surprises.

Chicony TK8880F / 2066A2

The board was examined in previous post at 160MHz.

--- 180MHz (3x60)

I shared already information about it running at 160MHz in this post.
It did well back then, so wanted to see how will manage the 180.

ADZ/ADW CPUs run well at default voltage.

Sees 64Mb EDO memory modules as 16Mb. Probably a BIOS issue. Already using the latest available BIOS, so no chance to improve on that. Used 2x64Mb (50ns) = 32Mb recognized RAM.

1024Kb L2 cache in WT mode. Not very picky about chips - used a mixture of 10/15ns ones.

This is hybrid PCI/VLB/ISA motherboard.
All PCI video cards i tried work well.
Tried bunch of VLB cards too Trident TGUI9440AGI, S3 Trio64, Ark1000VL, Paradise WD90C33 - none of them lit up. Orchid Kelvin 64-VLB (Cirrus Logic GD-=5434) worked, but had to significantly reduce L2 cache and DRAM timings for it to light up.

There is no on-board IDE controller. This caused a bit of trouble at 60MHz base frequency.
ISA IDE controllers worked right away, but performance was lacking.
Tried bunch of the VLB IDE controllers listed on page 5 in this thread - none of them worked.
Also, none of the Promise PCI IDE/RAID controllers worked either.
This is a motherboard specific issue, because they do fine in other systems running at 3x60, 3x66 or 4x50.
At the end, a Silicon Image UDMA66 PCI IDE controller worked fine.
If ISA IDE controller is used PCI IDE IRQ MAP must be set to ISA, otherwise to PCI.

The board struggled at 60MHz, so BIOS settings had to be lowered.
DRAM READ WAIT STATES = 1 (best is 0)
L2 CACHE WAIT STATES = 2-2-2-2 (best is 2-1-1-1)
IBC DEVESEL# DECODINOG = MEDIUM (best is FAST)
SLOW REFRESH = DISABLED
HOST-TO-PCI POST WRITE = 1 (best is 0)
HOST CLOCK/PCI CLOCK = 1:1/2 (best is 1:1)
Surprisingly this didn't hurt performance in interactive DOS graphics as much as i thought.
But here is the bummer - no matter what - Windows (both 3.1 and 95) refused to start on that rig - hung during logo screen.
Tried hard to overcome that - different CPUs and voltages, most conservative BIOS settings, different video and ide adapters, memory modules, L2 cache chips - no dice.
Also, SpeedSys hangs hard during initialization.

Shared are numbers for the interactive DOS graphics only since Windows and complex offline graphics tests are hard no-go.
benchmark results

Performance in DOS is pretty good, especially considering the inflated wait states listed above.
But the board is struggling in this configuration - too many issues.

--- 200MHz (3x66)

No lights entirely.

I've never been very good at the wick. Maybe I have junky stuff.

2-1-2 at 60 MHz would make for a nice Cyrix 5x86-120-S1R3 system if all the selected PCI cards worked at 60 MHz. If you have another rev.D, it would be interesting to see if 3-1-3 is completely stable with 15 ns. I suspect it is.

It is good to know that the rev.D w/10ns can now match the M918 in Quake (21.8 fps). I wouldn't want to build a system around an m918.

Which versions of the m919 are you testing?

Did you ever try to use the /D0:17 /F0 /SM0 options in the UMC v3.1 driver? I don't think the /SM0 option was in the readme. It improved my DOS performance on the UUD board, more than just using /D0:17 /F0. There is also /MM0 mode, or "MASTER" mode. Another user, jakethompson1, showed pretty good results with MASTER mode compared to /SM0 (standard mode) or /F0 /SM0 (standard mode with FIFO). Results from jakethompson1,

15,753 KB/s using BIOS PIO mode 4 setting and no driver
26,607 KB/s using UM8673.SYS /D0:17 /NF0 /SM0 
310,779 KB/s using UM8673.SYS /D0:17 /F0 /SM0
413,900 KB/s using UM8673.SYS /D0:16 /F0 /MM0

These results are with v3.20 and a magnetic HDD. The /NF0 forces FIFO disabled. These should be the buffered read speeds in speedsys. It is suspected that the linear read speeds may be incorrect. The /MM0 option, which may only work on v3.20, showed slower benchmarks on my system - perhaps due to single sector transfers (CF card) vs. multi-sector transfers on magnetic HDDs.

For my case, I have noticed consistently higher scores when using the XT-IDE BIOS compared to the MB's BIOS for onboard IDE. The increase isn't a lot, but it was repeatable - usually around 150 KB/s.

EDIT: I have tested the LS-486E rev. D w/10ns SRAM and 2x60 MHz using an IBM 5x86c. It cannot do 2-1-2. Best it can do is 3-1-2 (boot Win95c, run apps, run Quake). This is unfortunate, but not uncommon when on the verge of stability. The Am5x86, being a lot slower per clock, can often run stable with faster SRAM settings when teetering on the edge.

I agree the wick is much better for cleaning pads.
Lay your iron tip's flat side on the wick and once it gets up to temp just drag them together as one unit across the pads

maxtherabbit wrote on 2021-08-21, 13:05:

I agree the wick is much better for cleaning pads.
Lay your iron tip's flat side on the wick and once it gets up to temp just drag them together as one unit across the pads

This ^

Also, like everything else the desoldering braids vary in quality. The good ones absorb the tin instantly, even at gentle contact with it, others don't and act more like a grinder.
I use this one - works great.

---

I confirm LS D is completely stable with 15ns chips at 3-1-3 wait states. Can do 2-1-2 as well - fine for most DOS and Windows games/apps, but not completely stable.

Yes M918i is unsuitable for real computer at 180MHz.

M919 latest revision 3.4 b/f. Still joke of a system. 😉

Had an offline conversation with jakethompson1 about specific UMC IDE driver arguments, but my focus was on 3x60/66 + CF cards. Eventually got the driver to stop hanging at boot, but had to force-lower the settings to a point where it became much slower than no driver.
Yet to try at 4x40 and 4x50 with MM0 and F0, have it only with D17. Are your numbers with Cyrix or AMD processor ? Also what base frequency ?
Will give it a spin at 4x40/50 later today. Wanted to check ever since Jake mentioned it.

pshipkov wrote on 2021-08-21, 14:33:

Are your numbers with Cyrix or AMD processor ? Also what base frequency ?

If you are referring to the UMC HDD results, I didn't present my numbers, I presented jakethompson1's numbers, which were with a x5-133/4x. I was unable to replicate his numbers with an x5-133/4x. The difference in the setups was that I had a CF card in use while he used magnetic HDD.

Ah, yes.
These old controllers show better numbers in tests with mechanical HDDs than with CF cards.
In reality however CF is much faster than those old IDE HDDs, mostly attributed to seek time and random access.

Btw, been thinking about the video setup of the 180MHz PC i am starting to put together.
Bolted a Voodoo 1 in the Asus PVI based rig from page 1, but 486 CPUs are just insufficient to handle any 3D video games, so never use the Voodoo actually. It is kind of for optics only and few tests here and there. Same applies for any Rivas, TNTs, GeForces, ATI something-something ...
Also, there was no such thing as common graphics libraries for (486 class) PCs in the mid 90ies, so no application will take advantage of them either.
Thinking to go with video card such as Matrox Millennium or Millennium 2 that excels at interactive DOS graphics and Windows GUI. Probably with 6, or 8 Mb of video memory - for bragging rights.
Cirrus Logic GD-5480 gives small performance increase in DOS interactive graphics, but MGA is the better place overall.

Any thoughts ?
Feedback welcome.

I personally haven't noticed much of a difference in general use, like boot-up and game loading with CF vs. a late generation magnetic HDD. I do miss the mild magnetic HDD sound though because it reminds me the system is still working and hasn't hung up.

I've read on here that the Matrox G200 was also pretty good with DOS graphics as well. If you want bragging rights w.r.t. RAM size, get a G200 with 16 MB. Since you are building a hot rod 486, I think people might be interested in game benchmark results, so a Voodoo2 will offer the best 3D game results with less fuss. If wanting to be more period correct, then a Millennium 2 + Voodoo1 suits. If wanting to omit conventional 3D (DirectX/OGL), I think it might be fun to play with S3D titles, so perhaps a ViRGE GX. I recall reading somewhere that there may have been a ViRGE variant did 8 MB.

feipoa wrote on 2021-08-22, 04:44:

I personally haven't noticed much of a difference in general use, like boot-up and game loading with CF vs. a late generation magnetic HDD. I do miss the mild magnetic HDD sound though because it reminds me the system is still working and hasn't hung up.

With period correct HDDs there is a difference, mush less with late IDE HDDs.

feipoa wrote on 2021-08-22, 04:44:

I've read on here that the Matrox G200 was also pretty good with DOS graphics as well. If you want bragging rights w.r.t. RAM size, get a G200 with 16 MB. Since you are building a hot rod 486, I think people might be interested in game benchmark results, so a Voodoo2 will offer the best 3D game results with less fuss. If wanting to be more period correct, then a Millennium 2 + Voodoo1 suits.

G200 may be a good option too.

I am going with a custom made very compact PC case that wraps tightly around the LuckyStar D motherboard with just enough space for the right size cooler.
I was going with PCI cards that fit within the dimensions of the mobo itself, but your proposal for G200 creates a dilemma - G200 quad head for maximum pimp or MGA/MGA2/G200_single_head.
G200 Quad head is wider than LS D. Hmm ...

To illustrate the problem:

feipoa wrote on 2021-08-22, 04:44:

If wanting to omit conventional 3D (DirectX/OGL), I think it might be fun to play with S3D titles, so perhaps a ViRGE GX. I recall reading somewhere that there may have been a ViRGE variant did 8 MB.

Here is a post comparing couple of Virges. One of them is the 8Mb VX model.
Inclined to pass on Virge at the moment.
They are good 8 bit pixel pushers and decent at Win GUI, but for 3D ? Nah.

I'd size the case for the UUD in case you change your mind. Maybe you do the 1024K mod and find the UUD the better option, esp. if the IDE is faster given recent revelations.

I've had issues with the Quad G200. It wouldn't work on the UUD, for example, while the G200 worked.

Oh nice, you were able to find the memory upgrade card for that ViRGE VX! Those Those GLQuake results were horrific, even on the VX. No wonder they called S3D a decelerator.

One of my favourite combinations is the G200 w/Voodoo2. If you have a quality pass thorugh cable and a decent Voodoo2, the on-screen noise or distortion should be negligable. For added flavour, do SLI.

Initially I was thinking to size it to UUD, but then decided to go super compact. Reconsidering that already.

Didn't mess much with these later pci video cards. So good to know that g200 may cause trouble.

Haven't spend much time profiling voodoo cards.
Is there an advantage running voodoo 2 given the fact that the cpu cannot saturate even voodoo 1 ?

Yes I studied the Voodoo2 vs. Voodoo1 affect in depth on the UUD board with a Cyrix 5x86.

Actual side-by-side frame rates for Voodoo1 vs. Voodoo2 w/Cyrix 5x86 shown in this chart:
download/file.php?id=50985&mode=view

On this next chart we see the percent increase in performance of the Voodoo2 compared to the Voodoo1 w/Cyrix 5x86. The average boost w/Voodoo2 is around 15%, but in some games (Tomb Raider), it was as high as 36% faster w/Voodoo2, while in others (Descent2/DarkForces 2), the two yielded identical results. Of the 13 games benchmarked, 11 yielded better results with Voodoo2 compared to Voodoo1 on the UUD.
download/file.php?id=50986&mode=view

Thanks. I remember seeing these before.
Ok, so Descent is the only playable game.
Still, will probably add a Voodoo card for completeness.

BTW, do you have observations of voodoo 3 pci cards ? I think there is even voodoo4 pci.
I suspect they can get problematic on 486 hardware ?

There is some info online about them, but inconclusive.
Usually in the form of "hey voodooX card worked fine in my motherboardY" and no further detail provided.

What one determines as playable is subjective. It seems like more people were content with slower frame rates in the mid-90's. Of the games on that list, I find these perfectly playable on my IBM 5x86-133/2x with Voodoo2:

Descent 1
Descent 2
GLQuake
Tomb Raider
Tomb Raider Unfinished Business
Outlaws
Forsaken
Dark Forces 2

This list is not conclusive and only games with a frame rate counter are on this chart.

From your response, it looks like you consider 45 fps as the minimum requirement for 'playable'.

Voodoo3's and Banshee's should work on the SiS 496 motherboards. I have a DTK board w/SiS 496 and Voodoo3 setup. I've not run into any problems with this combination. UMC boards don't like Voodoo3/Banshees, but do well with Voodoo2's. I don't own Voodoo4/5 cards, so cannot comment, but another notes the Voodoo4 didn't work on his LS-486E.