Back then it was different.
For example ‐ remember playing Mortal Kombat 1 at 10 fps - basically a slide show - didn't make me upset at all.
I was just so fascinated and grateful that I had the opportunity to experience this at home.
Today – a bit more picky. The magic is no more, only memories of it. So, details like fps matter.

But also- I recognize that will probably fire couple of games on that retro PC for the sake of seeing them run on a 180MHz system, but unlikely to spend time playing them on it.
Because of that any fps will be fine really.

I am sparse on Voodoo cards. Not very interested in them in general.
Looked up some prices of voodoo4 - stoopid expensive.
Back then any voodoo card after 2 was questionable investment given what their competition was offering.
Probably that's why 3,4,5 are so scarce today = inflated price.

Didn't find conclusive information online about voodoo 4 and 486 class hardware, but I can see how they are apart enough to be trouble.
Given the limited space inside the compact PC case and the relatively large cpu cooling contraption - the PCI cards config will be influenced by their dimensions. Will need to get creative there.

Voodoo 3 was kinda spendy at release (Though pricing makes it look lower mid-end in todays market) and made the OG TNT look bad, a lot happened though and in a few months it was trading blows with the TNT2 which made it look more ordinary, then VSA-100 was announced and Radeon and Geforce came on the radar, and in early 2000 they dropped the price on the V3 significantly, then in just a few more months it found itself the entry level product against 4 and 5. By that time they had it selling not far over a hundred dollars, and 3DFX was grinding them out as fast as it could at the STB Mexico facility. Then after financial troubles and buyout there was still significant stock of V3 around for a couple of years, and it was still appearing on OEM boards, so chips must have been available too. So I think plenty of V3 were manufactured. With XP coming out it was the support thing that made them get retired earlier than TNT2 did probably. Meaning they got recycled or tossed a couple of years earlier than contemporary cards. I think really they're not really all that rare though, it's demand that makes them pricey now. V4s and V5s, much shorter manufacturing run, price stayed high while they were being manufactured, didn't sell apart from to Voodoo fans because of Radeon and GeForce. So those should be rare. V3s not so much, unless they got thrown out more and kept hold of less. There might be more V5s around than V4s, because voodoo fans wanted the ultimate voodoo, they never got cheap used, they were sought after from end of production. Maybe mid noughties they were under a hundred for a while, but that was a lot when used GF4 4200s were going for $20-$40 and you could probably get a Radeon 9700 for the same money as the V5s.

But anyhoo, threadjacking I guess... the FPS thing, I think it is changes in FPS that make things unpleasant to play, you can have 10fps solid, and get used to it and anticipate things.. if it's 15fps down to 3fps then 7, then up to 20 when you get in a tight space with nothing much to render, then it's super annoying and you can't judge your movements well. The average says nothing about this really, you'll get some games average 45 fps and you're looking to get 60 because it's all over the place but some do 30 and it's great, steady pace, no huge drops. Tomb raider for example, lot of cards/CPU you're running around the environment real smooth, then all of a sudden Bats! and it's a freaking slideshow.

Not a hijack. Appreciate the info.
Agreed with your comment about frame rate variability.

Looked at what late PCI video cards are in the bag.
It is not that much really (haven't been that interested in them i guess):
G200/16Mb
G200 Quad head + TV out
TNT2/M64
GF2 MX400 64Mb + TV out
Voodoo 1

Thought i have Voodoo2, but not really.
Looking at Voodoo3 2000/3000 prices - unlikely to spend hundreds of $ on them. Not convinced they do better than TNT2 to start with. (or am i wrong ?)
TNT2/16Mb is a bit more reasonably priced, but still $100++.

If the TNT2 and/or MX400 match MGA for 8-bit pixel pushing, that will pretty much decide it.
Otherwise it feels like MGA/G200 + voodoo 1/2.

If you don't care about 3D performance, then you can use any decent 2D card. If you are after best frame rates, then 3dfx glide has the least CPU overhead and will result in better frame rates. On a 486, the TNT, TNT2, GF2, and G200 offer far worse frame rates compared to 3dfx glide based cards.

The Voodoo3 was the last 3dfx card of interest to me because it works on a SiS 496 boards and offers space saving adtantages. It offers clarity and performance similar to a system with a G200+Voodoo2+Voodoo2 (SLI), so if you want to use PCI slots for ATA, SCSI or NIC, you can.

There have now been two reports indicating that Voodoo4/5 cards don't work on socket 3 (UMC & SiS) systems. I guess I should have bought more Voodoo3 PCI cards when they were cheap. For PCI, I think I only have one V3 3000 and one Banshee. Would have been useful to keep an extra around for bench top tests.

Thanks for the info.

Fastest possible accelerated 3d raphics will be a welcome bragging point, so it counts.
Going to spin those late pci cards I have here to feel them out, but it sounds like you walked that path already, so I trust you will.
Based on your notes g200+voodoo 1/2, or voodoo3 is the way to go. With bonus point for Voodoo3 for space saving.
Waiting on compact cpu cooling unit to come. Once I put it on, the pci cards config will become clear.

Do you have a link or spec sheet or URL for this "compact cpu cooling unit"? The use of the term "unit" makes it me think peltier, heatsink, and evaporator all in one?

There is no off-the-shelf hardware set for cooling 486 CPUs with Peltier element on top.
Where i am going currently:

About the main heatsink:
Too big for 5V Peltier element, yet to see if it can handle 12V - just in case.
This type is preferable than the classic flat ones (that sit directly on the CPU) because of two reasons:
- At 12V Peltier the flat ones are not good enough, which requires wider frames that don't fit in the compact motherboard layout.
- While bulky overall, the base is compact and matches the CPU dimensions, which is what matters.
This btw is the smallest heatsink of that type i was able to find online - probably not obvious from the screenshot.
Fan is too bulky and has funky candy color = no go. Will be replaced by slim and quiet Noctua model that proved itself in all the testing so far.

One or possibly two 20mm fans will blow at the CPU area to disperse the minor water condensation. Hope one will be enough.
Custom 3D printed frame will hold the whole thing together.

Probably the last (for now at least) addition to the 486 at 180MHz - an exercise in practicality investigation.

PC-Chips M919 V3.4B/F - last revision of a long line of infamous motherboards.
There is plenty of information online about it but not much about its performance qualities.
So, let's see how this guy does perf-wise.

Used this one for the tests:

There is another one still in the original box + cables, booklet and Am5x86 133-ADW CPU.

Supports all Socket 3 CPUs of its time. A bit convoluted jumper setup.
ADZ CPU models require 4V, while ADW do just fine with 3.3V.
Noticed that some motherboards have such "requirements", while others handle both CPU types at 3.3/3.45 no problem.

UMC UM8881F-EYT/UM8886BF-FXT chipset.

Custom 256Kb L2 cache module with 15ns rated chips. Frees space on the PCB, but upgrades are difficult.
L2 cache is used in WT mode. Performance is same or faster than WB, depends on which test you ask. This is the norm for most 486 motherboards. Only some rare exemplars such as Asus PVI handle WB properly.

Supports EDO RAM.
For some reason performance suffers if exactly 64Mb memory is installed, but fine with any other number. This is confirmed by other users later in the thread.
Tested with 32Mb 60ns module. (EDIT: 2 years later - additional testing with 128Mb was performed as well while assisting somebody to figure out issues with his M919.)

Favors PCI over VLB which is in line with most PCI/VLB hybrid implementations based on that UMC chipset. Didn't like ARK1000VL VLB video cards (best for DOS graphics). Used S3 Trio64 VLB instead, which is a hair behind in DOS, but does way better in Windows GUI.

Implicit bus divider. FSB over 40MHz causes it to automatically kick-in. This is remedied easily by POSTing with TURBO turned-off. Still annoying.

Latest BIOS 09/15/1998S.

The built-in IDE controller is not great.
Also, can hang sometimes (very rarely) upon soft reset with CF cards bigger than 8Gb. Hard reset fixes it.
Didn’t encounter this problem with smaller CF cards.

--- 200MHz (4x50 or 3x66)

Cannot get the system to DOS prompt reliably.
Tried hard with 2 different motherboards, different processors, memory modules, with and without L2 cache, different CPU voltages.
Feels like it is something on motherboard level.

--- 180MHz

All BIOS settings on max, but for complete stability the next 2 exceptions are needed:
CACHE SPEED OPTIONS 3-2-3 (best is 2-1-2)
DRAM READ WAIT STATE = 1 (best is 0)
The system can operate with L2 cache set to 2-1-2 and achieve better numbers, but is not stable.
This was already showcased in another thread by user CPUGalaxy.
Considered to eventually upgrade the cache module with 12 or 10 ns chips, especially if all other aspects of the boards were top notch ... but they were not, so didn't bother at the end.
Still, may get to it at some point later and update the post.

UMC IDE driver works with speed set to 17 (max is 18, undocumented) and MM0.

The usual set of benchmarks, compared to LuckyStar D and Biostar UUD.
Provided are results with both MGA PCI and S3 Trio64 VLB graphics adapters.

Obviously not the fastest motherboard out there.
Comes with too much baggage - implicit bus divider, limited RAM (or things get slow), IDE issues, basically non-upgradable L2 cache.

1024Kb L2 cache
Managed to obtain 1024Kb L2 cache module from a fellow Vogon user (he produced them).
It turns M919 into what looks like the best motherboard at 180MHz.

All BIOS settings on max except:
DRAM READ WAIT STATE = 1 (0 is best)
Basically, the faster rated chips of the new L2 cache module allow for CACHE SPEED OPTIONS to be set to 2-1-2.

Very satisfying.
System is fully stable.
The fastest overall motherboard in the 180MHz category. Effortless. Just works.
Attention must be given to the automatic PCI bus divider - boot in 40MHz FSB then switch to 60MHz (pull one jumper).
The built-in IDE controller is a bit lacking compared to the other qualities of the system.

--- 160MHz with 1024Kb level 2 cache

All BIOS settings on max. Fully stable.

Speedsys (this is an old screenshot with 256Kb level 2 module):

benchmark results

While performance gets a nice boost, the larger L2 cache buffer is still not enough for the board to outdo the rest.

Buffered read is only 5736 KB/s with /D0:17 /F0 /MMO ? Seems kind of slow doesn't it?

Funny that your v3.4B/F has only one BJT transistor for the VRM, whereas my v3.4B/F boards have two. I guess PC Chips threw out PCB revision control too!

QA guy turns it on. Board lights up - initial screen.
QA lead: Good job. Ship it.
QA guy: Hey, something may be missing here ...
QA lead: I said SHIP IT !
QA guy: Ok, ok.

Joking aside, I noticed that there are versions with two VRMs. The one i have works fine with dx2, dx4/5 and pod83/100 processors. Also, can do 200mhz dx5 at 4v. So all is good it seems.
So, not sure what the role of the second vrm is. Tracing the lines may reveal its purpose.

Yes, IDE perf is not shockingly good, but the driver does something, because the linear read speed goes through the roof with it.

Should compare part numbers, the whole string, might have like an XXX-C part that handles 20Amps and they couldn't guarantee supply, so some boards got XXX-B parts in tandem that only handle 12A each or something.

Read me the part number on your BJT and I'll read you mine. I bet they are the same, if there is even a part number. They probably ran out of this part on that particular day and decided to ship it anyway.

With respect to IDE speed, did you do a timed data transfer test to ensure that the "linear read speed" is correct? It is possible that only the buffered read speed is precise. I had some "linear read speeds" approach 30,000 KB/s, which is way out of spec. for PIO-4.

pshipkov wrote on 2021-08-30, 19:45:

UMC IDE driver works with speed set to 17 (max is 18, undocumented) and MM0.

feipoa wrote on 2021-08-30, 20:54:

With respect to IDE speed, did you do a timed data transfer test to ensure that the "linear read speed" is correct? It is possible that only the buffered read speed is precise. I had some "linear read speeds" approach 30,000 KB/s, which is way out of spec. for PIO-4.

It's the same issue we were running into. The linear read speed is wrong; the buffered read speed is correct. My suspicion is that something about the bus mastering feature in UM8673.SYS (which is what /MM0 does) also messes up Speedsys' sense of time. Perhaps it keeps interrupts disabled for too long. Passing /SM0 will put the driver in FIFO mode (no bus mastering) and will give a more correct (possibly somewhat slower) speed.

The slow speed is likely because only mechanical HDDs seem to support Read Multiple Sector commands aka block mode for PIO mode. I get 13,900 KB/s buffered read speed with a 40GB IDE mechanical drive and UM8673.SYS /D0:16 /F0 /MM0. Presumably that is as fast as the UM8886BF is capable of.

It would be worth starting a new thread just for UM8886AF/UM8886BF IDE issues as, as we have talked about privately, there is a lot there.

VRM model is C3420.
Datasheet: https://www.jbprojects.net/projects/wifirobot/c3420.pdf

Didn't time the IDE speed the old fashioned way. The overall performance and stability didn't meet the expectations, so spending more time on IDE itself was pointless.

Agreed Jake.
I noticed that the UMC IDE driver can confuse SpeedSys for good. But checking with ATTO for example shows believable numbers. So far didn't recognize specific pattern, but i think you are right. SM0 keeps the numbers real.
Confirm from my side that the UMC IDE controller/driver show higher perf numbers with mechanical HDDs compared to CF card, but CF are similar, or faster in the most cases. Mostly attributed to seek time / random access.
It will be good to dump the collected data into its own thread.
Get some numbers organized by mobo, BIOS settings, storage device, driver settings. So we can see it.
I will be folding the 180MHz circus tent soon and will participate.

I'm still puzzled by your slow buffered speedsys result. My best result with a UMC IDE controller is with v3.1 using /D0:17 /F0 /SM0 at 8327 KB/s. Speed 16 and 17 yield the same benchmark results. Of course, that is with CF cards 8 GB or less.v3.2 of the driver yielded 7453 KB/s.

lol, yes, someone should start and compile the UMC driver data onto a new thread. Obviously, when someone (me) says "someone" in this context, it means "please not me".

on one of my flakey m919 v3.4B/F boards, there are two of these BJTs: D2583

on my other flakey m919 v3.4B/F boards, there are two of these BJTs: C3420

So looks like with your batch of these boards, they ran out of parts to meet their daily or weekly quota.

If you are talking about the M919 IDE performance - i assume you compare it to Biostar UUD.
It that's the case - that's apples and pears. M919 << UUD.

I doubt PC-Chips were soldering unneeded elements on some of the mobos.
At the same time only one VRM works just fine. Btw, tried to trace the left VRM lines, but they seem to be in the middle PCB layers. Only one fat line goes to a capacitor, the rest is not visible.
More riddles.

Yes, referring to UUD's performance. Is there such a large divergence in UM8886BF performance on the m919 using the same IDE controller vs. uud?

About integrated IDE controllers:

Looking at the UMC numbers, they vary from 5 to 7 mb/s at 180 mhz.
But most of the time without driver, since it tends to hang the system at that frequency.
At 160 mhz speed is between 4 and 7-8 mb/s, with driver.
That is with CF cards. With HDDs - numbers go up, but still a wide range between best/worst.
Overall integration and optimization of the motherboard matter, obviously.
UUD seems to be the best - between upper 7 to lower 8 mb/s.

SiS boards are like that too.
From 4 to 9 mb/s, with Asus PVI leading the pack with ~9.5 at 160mhz.

Two updates before forget the details, especially about the second one - it may be useful to others one day.

---

Found a Voodoo3 PCI graphics adapter for $125 including shipping and tax.
Initial ebay survey was chilling, so i discarded the idea of using 3D accellerators in the 180MHz rig and considered 2D pixel pushers only, but out of nowhere V3 shown-up on the radar.
Compared to the silly prices scalpers ask online the hundred dollars spent don't sound too bad.

I have been fine-profiling the LuckyStar revision D and Biostar UUD motherboards and was still not clear which one will end-up being the better overall option.
But the Voodoo3 tilted things in favor of LuckyStar D, since Biostar UUD does not light-up with the card plugged-in.
V3 came with 6ms memory (166MHz rated).
In 160MHz system (4x40) it overclocks just fine to 190-200 MHz. Which is basically Voodoo3 3000.
In 180MHz system (3x60) it can go no higher than 165-175 MHz.
But 3D graphics performance is entirely CPU bound, so video card overclocking delivers the massive performance increase of 0.1 fps at 640x480 and no difference at 320x240.

Captured some action on video.
Over 32 fps in GlQuake on 486 junk. 😀

L2 cache timings are now at 3-1-2 (previously 2-1-2).
The more granular testing indicated that 2-1-2 used in the previous coarse tests is too ambitious.
Increased wait states was expected before the final assembly.
So far things behave stably (still inconclusive, since didn't test with lan and sound cards plugged-in).
Will see where this one ends.

---

The most crucial aspect of this 180MHz business is the CPU cooling.
Spent some time to profile and better understand the dynamics at play.
AMD 486DX5 ADZ/ADW CPUs with good dies can operate reliably at 180MHz when surface temperature is around 24-26 C.
These old processors maintain almost identical thermal profile regardless if they are under load, or idle - this simplifies the problem i need to solve.
Air cooling does not guarantee temperature range. It heavily depends on the ambient temperature.
So, for instance, system can be stable in the morning / evening, but not at noon.
Using Peltier element ensures much better temperature consistency and tighter range.
The problem is that at 5V there is water condensation buildup.
Until now i was thinking of some elaborate contraption with big heatsink+fan and additional fans blowing at the CPU area.
But initial tests were inconclusive this will actually work, which made me reconsider the approach.
There are several tightly coupled factors that contribute to the thermal stability and range:
- fan size
- fan speed
- heatsink size/shape
- heatsink material
- voltage to Peltier

Started with dropping the voltage to the Peltier element. Played with resistors in the range of 20-30 Ohm.
Quickly realized that it is better to keep voltage at 5 and only adjust heatsink/fan = one variable less to balance.
After some experimentation with different components, this is what ended-up working well:

Relatively small heatsink from Asus motherboard. This one has fewer ticker ribs which reduce its efficiency compared to another heatsink with the same dimensions and more + tinned ribs,, which cause temperature to drop too much. Smaller heatsinks than this size overheat, regardless of how big fan blows at them.
5V to 50x50mm fan.
Test assembly. Needs to be improved. Will most likely require 3D printed holder.
Ensures 20-23 C on average. No water condensation. Simple and compact.

Then thought that temperature conditions will most likely be different inside a closed PC case.
Tested for quite some time like this:

Good that i tried it. Things were much hotter than expected.
Luckily bumping the fan to 12V brought back the desired equilibrium.

It may require additional tweaks before the end, but so far the path ahead is clear.

One last, but important note - Peltier element draws a lot of power. Using less than 450W power supply results in overheating 5/12V rails.

Are these dollar store peltiers temperature adjustable by varying the voltage, e.g. run a 12 Vdc peltier at 5 Vdc or 9 Vdc? I haven't played with mine yet so I am perhaps asking an obvious question. If they are not generally adjustable in this fashion, have you plotted any data points to create a temperature vs. voltage graph at various ambient temperatures?