feipoa wrote on 2021-08-02, 22:09:

The PCI clock multiplier is not an undocumented setting. It is in the official manual, Chapter 2. Installation, top of the page […]
Show full quote

maxtherabbit wrote on 2021-08-02, 21:51:

Possible that we do have it mixed up, there is no pin numbering on the silkscreen and it's an undocumented setting anyway.
I'm fairly sure that the position closest to the card slot brackets is 1/2 and the position closer to the front of the case is 1:1.

What about your video card? Many of them do bus master, the Matrox Millenium that pshipkov and myself use notably does not.

The PCI clock multiplier is not an undocumented setting. It is in the official manual, Chapter 2. Installation, top of the page, above the REMARKS section. For whatever reason, they did not place this on the page with all the other jumper settings. The official layout of the board, in the manual, shows pin 1 being closest to the CPU. Jumper on pin 1-2 is for the 1/2 multiplier; pin 2-3 is for no multiplier.

I have attached the official manual for your reference. Go to PDF page 10.
LuckyStar_LS486E_RevD.pdf

I have used a few graphics cards during these tests, but when the Promise Ultra100 was installed, I was using a noname S3 Trio64V2/DX because this card is small enough to fit in the first PCI slot. Does this graphics card bus master? There is a jumper installed on the card, JP1 on pins 1-2. The other options is 2-3. I have no idea what this jumper is for. With ISA cards, such jumpers were usually used to set IRQ to 9. ALso to set 0 WS / no 0 WS.

Indeed you're right - it is documented for the Rev D. It is not however on the C2. Also the jumper on the C2 is roughly in the same location but it is perpendicular to the Rev D's equivalent.

As for the video card bus mastering I don't know about the trio specifically. You can use the "PCI information" feature of SiSoft Sandra 2001 (which runs on Win98) to read out the config flags for all PCI devices in your system and see if it supports bus mastering or not.

Well, that's a sledge hammer of heat gun you got.

Below is what i use - mr. El Cheapo.
It has 2-3 more ends, that i couldn't find for the photo session - they are somewhere in the drawer.
The smallest one is half of that on the picture.
But 5mm (on picture) seems to work best, so it is used exclusively.

I do the job as shown in the video Max provided.
Except, i don't cover the whole board with slime.

Looks like you have a hot air soldering iron. I did refer to mine as a "hair dryer" style hot air gun, not a hot air solder iron. Does your el cheapo have a brand and model? I don't really want a proper and expensive hot air soldering station.

I seriously doubt that trio is setup in bus master mode, but who knows. I'm moving on.

I had my doubts about that copper wire trick at first, just like you, "wtf, lol, omg" etc. but I suspect it will work just fine if not left on for too long.

EDIT:
How about that piece of junk mentioned in the video posted by maxrabbit? Worth the $70 CAD ? https://www.amazon.ca/Rework-Solder-Station-S … 28041313&sr=8-5

EDIT2:
I'm watching this review. https://www.youtube.com/watch?v=vva2t21sOAs Love this guys enthusiasm! I bet he doesn't have kids with enthusiasm like this. He ends with "for the price you can't go wrong".

Yes it's worth it. A hot air station is pretty simple device, so I don't feel bad getting a "cheap" one. Conversely, it's also an essential tool in the electronics tech aresenal.

I don't blame you for moving on, but I would urge you to revisit the bus master burst setting. Disabling it cut my promise ultra133 transfer rates by almost 1/3

There's a brand called FEITA 858D and its only $45 CAD. Looks identical. Taking the family out to diner costs way more than that. I think I'm going to go for it.

model says gj-8018lcd
$50 today it seems.
but i am not sure if i am ready to recommend it.
if i keep it at 900C for few minutes - termal protection kicks in and it needs to go all the way down to below 100C before can be used again. That process takes another few minutes.
Never bothered me really and dont know if other heat guns are better in that regard, so choose wisely.
i think this part will make the biggest difference in user comfort if the tools is used often.

The copper wire extension apparently works too and if you dont go super high temperature on the soldering iron should be safe for the PCB for quite some time.

Dave is a giant among electronics youtubers, I'm shocked you're not familiar with him. He is a dad btw

pshipkov wrote on 2021-08-04, 02:29:

model says gj-8018lcd $50 today it seems. but i am not sure if i am ready to recommend it. if i keep it at 900C for few minutes […]
Show full quote

model says gj-8018lcd
$50 today it seems.
but i am not sure if i am ready to recommend it.
if i keep it at 900C for few minutes - termal protection kicks in and it needs to go all the way down to below 100C before can be used again. That process takes another few minutes.
Never bothered me really and dont know if other heat guns are better in that regard, so choose wisely.
i think this part will make the biggest difference in user comfort if the tools is used often.

The copper wire extension apparently works too and if you dont go super high temperature on the soldering iron should be safe for the PCB for quite some time.

900C is outrageous, I keep mine at 380C

I've seen a few of his videos now, but not enough to know his name and children. I've always liked his enthusiasm. Let's see if he keeps up the enthusiasm a few children later. He might be safe if he has a more traditional partner (think 1950's homemaker) who does most of the child rearing.

I tend not to watch too much youtube and prefer forums and other static online content. Seems that times have changed. If its not on youtube, it must not have happened.

900 C? I think he meant F.

Actually i meant Celsius. 😀
My workflow is one multi-pin component at a time as quickly as possible. This implies higher temperatures. I usually go for 700-800 range, but the controls on the heat gun are not great - need to balance between airflow and temperature. I guess i should have the settings figured out already, but don't really heatgun often enough, so the next time is always a bit of guess work and if don't pay attention can easily overshoot = forced cooldown after a while.

As for dads and kids.
Trick is - parents to not get outnumbered.
Past that point things can get out of hand.

Have you checked to ensure that your el crapo hot air gun is actually able to output 900 C? These cheap hot air guns don't normally go above 500 C. What would be the need for 900 C? That's 1652 F.

I think even if you had just one kid and and he/she is moderately autistic with ADD, ADHD, oppositional defiance disorder, you'd be "out numbered". I'm obviously describing my middle of 3 children. He also his this compulsive disorder to break any and everything. If it can come off, he will take it off, even if it is breaking it off. Just as I'm typing here, he broke our bedroom's floor aircon. Another tough condition is if they are schizophrenic without meds (they nearly always stop taking their meds!), that's another situation for being out numbered with one kid. A good book on that subject is Hidden Valley Road: Inside the Mind of an American Family.

I think a good case for having children is to have a stay-at-home wife/maid who's libido isn't out of control. I think it may be possible to tie modern day divorce rate increases with the increase of both parents working full-time. I think the military may still have a high rate of homemakers, but that has probably been changing recently too. My advice is just not to have any.

Well, I obviously never read the manual.
Just checked the temperature on the output. It is 400C when heater is on max. The digital display on the device is in F.
So in other words, this heat gun can operate for long periods of time when around 350-380 C. Above that it will eventually shutdown itself.
Somehow I ended up using the device without sweating comprehension of the displayed numbers - instead going by gut feeling - like a "pro". 😀

These conditions make things extra challenging.
Social and economic reality definitely have an impact on families.

feipoa wrote on 2021-08-04, 04:46:

I'm obviously describing my middle of 3 children.

I'm really sorry to hear that bro

feipoa wrote on 2021-08-04, 04:46:

I think a good case for having children is to have a stay-at-home wife/maid who's libido isn't out of control. I think it may be possible to tie modern day divorce rate increases with the increase of both parents working full-time.

based!

Picking up where i left the 486 at 180MHz - an exercise in practicality several posts ago.
This time with PC-Chips M918i revision 1.2.

It was not clear if the on-board clock generator actually supports frequencies above 50MHz.
First i thought that the clockgen is relabeled component, but there are online stores still selling it, so it is legitimate after all.
At the same time, there is no trace about its datasheet online.
Had to verifying carefully if it can do 60/66MHz using trial and error.
Speed selection is a bit weird - 4 jumpers. That's quite a few configurations which suggests wide range of supported frequencies.
Tested multiple times with trusted components - 60ns FPM and 50ns EDO memory, 3 sets of L2 cache chips - 256Kb/512Kb/1024Kb configurations, ADZ/ADW CPUs that handle 200Mhz just fine.
Intention was to minimize the chance of missing the desired frequency because one of the components failed at it and the system didn't lit up.
Unfortunately the only supported configurations are 25, 33, 40, 50.
No 180MHz for this board. Too bad.

To compensate for my disappointment decided to check how it does at 160MHz (4x40) and 200MHz (4x50).

Replaced the flaky L2 cache the board came with (pictured above) with trusted ones.
At 160MHz 1024Kb is achievable.
EDIT: At 200MHz the best i could do was 512Kb single banked, but still interleaved (?). <- Cache is not interleaved. Confirmed in following posts.
There is something really strange about this chipset.
Its direct memory access is very fast. In fact it is so good that L2 cache can easily be slower than that. Can this be some in-chispet caching mechanism ?!
Additional details here.

Board likes EDO memory better than FPM. Performance noticeably improves with EDO.
If more than 32Mb RAM installed - performance tanks.
Tried to figure out why, but couldn't - feels like it is related to whatever internal caching is going on - probably limited cache size.
I saw similar issues in M919 and some other M boards.
It is weird. Any hints ?!

LBA mode has to be turned off, or the board cannot boot from some CF cards.

BIOS shows only up to 256Kb L2 cache in the boot screen.
Also, when running 4x50 it shows 150MHz.

--- 160MHz:

I usually start testing with interactive DOS graphics. They go quick and are great for coarse perf tuning.
With all BIOS settings on max the system could cycle in Quake, Doom, Check-It, etc., forever and ever.
Same for the round of Windows tests.
Great !
Then onto 3D rendering test. Very sensitive - great for fine perf tuning.
Instacrash.
Had to relax some of the BIOS settings to pass that stage:
DRAM WRITE TIMING = FASTER (best is FASTEST)
SRAM READ TIMING = 3-2-2-2 (best is 2-1-1-1) <- ugh, the used L2 chips can do 2-1-1-1 in other boards, this is a M918i specific issue
SRAM WRITE TIMING = 1 (best is 0)

As a result performance went down quite a bit.

benchmark results

--- 200MHz

I will be checking the 200MHz more thoroughly at some point later, so won't go into details here.
Already shared some early findings in this post.
The highest Quake 1 score so far.
There is implicit bus divider that kicks-in at 50MHz, but the system can take 1:1 no problem - just boot with turbo off, then turn it back on.
In general, things felt stable, but will see what this really means given the 160MHz experience.

Something curious:
Despite 12V Peltier in use to freeze the processor:
M918i refuses to boot with ADW CPUs that are perfectly capable of 200MHz, regardless of what the voltage is.
At the same time ADZ CPUs happily run at 4V on it.
Go figure.

EDIT: 64Mb RAM with less 512Kb L2 cache is unstable requires DRAM WRITE TIMING set to FASTER, which affects performance.
The simple cure is to step down to 32Mb.

---

I don't know if the above notes convey it, but this motherboard is one big contradiction.
It has its nice sides, but not enough to outweigh all the bad baggage.
Some 486 era PC-Chips experience at play.

I'm a little confused by these statements, which on the surface seem contradictory.

"Double banked, interleaved L2 cache running at optimal BIOS settings only appears to match direct RAM access in SpeedSys (picture below in this post), but actually has positive performance impact."

then in the reference link you posted

"Direct RAM access is actually pretty good. Single banked L2 cache at 2-1-1-1 only matches it, while 3-1-1-1 for double banked has negative performance net effect."

In the first sentance, you said that double-banked can run at optimal BIOS settings (presumabley 2-1-1-1), but in the other link, you say it runs at 3-1-1-1. Could you clarify this?

Yes, past reports with the FinALi show it having noticably better EDO performance, more so than other PCI 486 boards.

When I tested the m918 many years ago I also noticed the very fast RAM performance and that L2 couldn't increase upon it. The only way I could determine that L2 was actually working was to add many wait states to the memory, then I'd see the expected drops in Speedsys.

Does LBA mode still need to be disabled with certain CF cards, if for example, you set mode PIO-3 ?

How does performance tank with more than 32 MB RAM? Are you implying that you have to increase memory wait states to make it stable, or some funny kind of memory-not-cached thing going on? Or some automatic wait states being added to read/write? What are the metrics on the decrease in performance, e.g. using Quake 1.06 and cachechk memory/L2 read times? Is there an L2: WB/WT mode you can set? What if you physically remove L2 entirely, does the performance still dive after 32 MB is installed? Do you have other FinALi boards you can contrast this phenomenum with?

The FinALi M1487/M1489 chipset brief shows it suppors 128 MB EDO, up to 1M SRAM, 8bit tag, always dirty, or 7 tag bits, 1 dirty.

I checked my notes and I don't have any FSB specs noted above 50 Mhz.

FYI, from what I can recall, these boards often came with fake cache chips or no cache chips.

That's a neat trick to get 50 MHz in the PCI bus. Any idea if that also works with the m919? In the past I've wired the FSB jumpers up to the turbo to switch from 33 Mhz to 40 Mhz once it booted. This would get my PCI back to 40 MHz, but if your trick works, that is even better.

You can't boot ADW CPUs? What voltage is the board outputting on its 3.3/3.45 V setting? I've seen a board or two want 3.6V for Am5x86 CPUs.

Have you tried a Gigabyte GA-5486AL? It is another FinALi board I have, but it only takes single-banked SRAM. However, if I have interpretted your notes corectly, single-banked on this board can work with faster timings?

feipoa wrote on 2021-08-07, 04:53:

I'm a little confused by these statements, which on the surface seem contradictory. […]
Show full quote

I'm a little confused by these statements, which on the surface seem contradictory.

"Double banked, interleaved L2 cache running at optimal BIOS settings only appears to match direct RAM access in SpeedSys (picture below in this post), but actually has positive performance impact."

then in the reference link you posted

"Direct RAM access is actually pretty good. Single banked L2 cache at 2-1-1-1 only matches it, while 3-1-1-1 for double banked has negative performance net effect."

In the first sentance, you said that double-banked can run at optimal BIOS settings (presumabley 2-1-1-1), but in the other link, you say it runs at 3-1-1-1. Could you clarify this?

Good catch. Fixed it.
To clarify:
256Kb double banked, interleaved, 3-1-1-1 (was not able to get 9 chips stable at tighter timings, 1024Kb were even more unstable)
EDIT: 512Kb single banked, interleaved (?), 2-1-1-1

feipoa wrote on 2021-08-07, 04:53:

Yes, past reports with the FinALi show it having noticably better EDO performance, more so than other PCI 486 boards.

When I tested the m918 many years ago I also noticed the very fast RAM performance and that L2 couldn't increase upon it. The only way I could determine that L2 was actually working was to add many wait states to the memory, then I'd see the expected drops in Speedsys.

Yep.

feipoa wrote on 2021-08-07, 04:53:

Does LBA mode still need to be disabled with certain CF cards, if for example, you set mode PIO-3 ?

What i see here:
Transcend CF cards = good.
SanDisk Extreme CF cards = no-go if the option is activated.

feipoa wrote on 2021-08-07, 04:53:

How does performance tank with more than 32 MB RAM? Are you implying that you have to increase memory wait states to make it stable, or some funny kind of memory-not-cached thing going on? Or some automatic wait states being added to read/write? What are the metrics on the decrease in performance, e.g. using Quake 1.06 and cachechk memory/L2 read times? Is there an L2: WB/WT mode you can set? What if you physically remove L2 entirely, does the performance still dive after 32 MB is installed? Do you have other FinALi boards you can contrast this phenomenum with?

The FinALi M1487/M1489 chipset brief shows it suppors 128 MB EDO, up to 1M SRAM, 8bit tag, always dirty, or 7 tag bits, 1 dirty.

If more than 32Mb RAM performance just magically drops.
No settings changes.
Really puzzling.
I see the same in two other M motherboards.
Will capture some more detailed metrics tomorrow. Forgot the details.

feipoa wrote on 2021-08-07, 04:53:

That's a neat trick to get 50 MHz in the PCI bus. Any idea if that also works with the m919? In the past I've wired the FSB jumpers up to the turbo to switch from 33 Mhz to 40 Mhz once it booted. This would get my PCI back to 40 MHz, but if your trick works, that is even better.

Yeah, you can do it through the clock-gen jumpers, or simply toggle the turbo mode.
Turbo is much easier. Ask lazy to teach you the easy ways. 😉

feipoa wrote on 2021-08-07, 04:53:

You can't boot ADW CPUs? What voltage is the board outputting on its 3.3/3.45 V setting? I've seen a board or two want 3.6V for Am5x86 CPUs.

Board is doing 3.3/4/5.
Measurements confirm that.
No idea why ADW and M918i didn't click at 200MHz. Tried all voltages - no go. ADZ was just fine at 4V.
Obviously at 160MHz both CPU models were fine.

feipoa wrote on 2021-08-07, 04:53:

Have you tried a Gigabyte GA-5486AL? It is another FinALi board I have, but it only takes single-banked SRAM. However, if I have interpretted your notes corectly, single-banked on this board can work with faster timings?

EDIT: My understanding is that what matters is cache interleaving, but not the number of banks filled. Incorrect. L2 cache interleaving requires 2 banks.
I think it goes like this for 386/486 L2 cache:

1SIZE       BANK 0      BANK 1
2128K      32K x 8               <- non interleaved
3256K      32K x 8     32K x 8   <- interleaved
4256K      64K x 8               <- non interleaved
5512K      64K x 8     64K x 8   <- interleaved
6512K     128K x 8               <-[s] interleaved (? - some boards seem to treat this differently)[/s] non-interleaved
71024K    128K x 8    128K x 8   <- interleaved (This also works with 64K x 8 tag)

Do you know ?

At 160MHz i used line 7.
At 200MHz - line 2 and 6, where line 2 required 3-1-1-1 timings.

So this GA motherboard may be a better implementation of the ALI chipset.
I have Abit PB4 which works pretty well at 160MHz, but was not able to run it at 200MHz, despite its clock gen supports that. Wanted to give it one more try before give up.

So, what's your experience with M918i ?
Revision 1.2 ?

Did you see the same memory size related performance issues ?
Were you able to overcome them somehow ?

I'm pretty sure you need two SRAM banks to interleave the cache. Why do you beleive the FinAli can interleave with a single bank? I suspect the ALi chipset doesn't properly interleave with two banks (a bug?) which is why single-banked non-interleaved cache works with faster timings. But on the other hand, the fact that you could use 2-1-1-1 with 512K single-banked, but only 3-1-1-1 with 128K single-banked (line 2) is suspicious. Maybe 128K has a bug too. Are you able to use 2-1-1-1 with 256K single-banked (line 4)? It is also a little curious that nearly all FinALi boards only have sockets for a single bank.

I did not spend a lot of time with the m918 and GA-5486AL. I noted the discrepancy, was appalled, and moved on. I did notice that EDO worked, which was nice, but not enough to maintain my interest. I do plan on revisiting these two boards though. This is part of the reason why I don't sell hardware - something unique like that Quake result might turn up.

For your Am5x86-200 run at 21.8 fps - if you remove the SRAM entirely, is the score the same?

M918i and RAM amount. Verified very carefully.

160MHz with all BIOS settings on max, as described in the post above.

Here is what is going on:

Performance is consistent with 8, 16, 32 and 64 and 128 Mb. Didn't bother with 256Mb.
Results match the SpeedSys screenshot i posted above. Memory speed varies slightly between different runs - in the range of 237.5-239.5 Mb/s.

When 64Mb or more system memory installed - DRAM WRITE TIMING must be set to FASTER (best is FASTEST), or the system gets very unstable - often cannot complete POST.
This is a motherboard issue. The used 50ns EDO memory modules can handle much more extreme conditions.

Under 64Mb memory - DRAM WRITE TIMING parameter can be set to FASTEST.
The system is stable in DOS interactive graphics and the plethora of simple DOS testing utilities.
Stable in Windows as well.
Cannot pass the more sensitive tests, as described in the above post.

---

L2 cache. Verified very carefully.

Here is what is going on:

160Mhz, all BIOS timings on max, as described in the post above.

L2 cache is not fake.
Board does not like every L2 chip to post properly, but is not too picky. Still, the set of chips has to be curated.
Changing the L2 cache settings in BIOS has impact on stability. That's for sure.

256Kb in 2 banks is slower than direct memory access.
Double checked it by enabling/disabling EXTERNAL CACHE in BIOS = performance metrics change.
Faster when EXTERNAL CACHE is OFF, slower when ON.
Remember the inverted SpeedSys L2 graph from this post.

512Kb in 1 bank is on par with direct memory access.
Enabling/disabling EXTERNAL CACHE in BIOS results in no change in performance.

1024Kb in 2 banks is slightly faster than direct memory access.
Enabling/disabling EXTERNAL CACHE in BIOS results in minor performance changes.

feipoa wrote on 2021-08-07, 09:00:

I'm pretty sure you need two SRAM banks to interleave the cache.

Yes.
But for a moment this M918i L2 cache weirdness made me wonder if there is more to it.
The above clarifications cleared up any doubt i had.
Edited my previous posts - scratched-through the "maybe interleaved" inline questions. To minimize confusion for anyone else who may hit these posts eventually.

---

In addition - fired-up the Abit PB4 mobo from this post - wanted to make sure that i didn't miss something back then.
After all this M918i nonsense i started doubting my previous examination of the Abit board.
Fortunately the system behaves as a normal 486 motherboard, following all orthodox rules.

---

Not crazy about M918i.
It can do some nice tricks, but way too brittle and weird.

So, Feipoa, can you give the Gigabyte ALI board a spin ?
It will be interesting to see what it shows, assuming the vastly different implementations of the ALI chipset in PC-Chips M981i and Abit PB4.

256Kb in 2 banks is slower than direct memory access.
512Kb in 1 bank is on par with direct memory access.
1024Kb in 2 banks is slightly faster than direct memory access.

jeez, what's going on here? If you aren't changing the SRAM timings, there must be some automatic wait states added. Do you have the speedsys graph with 1024K? Is 512K in 2 banks also on par with direct memory access?

Unfortunately, my temporary and awkward motherboard test station is occupied at the moment. Doing XTIDE stuff then I'm returning to the LS-486E rev.D when my $38 hot air station arrives to swap the SRAM.