pshipkov wrote on 2021-04-04, 21:42:

Something weird. I sharpened this Asus VLI based PC for top DOS graphics performance. Recently it got some POD100 love as explai […]
Show full quote

Something weird.
I sharpened this Asus VLI based PC for top DOS graphics performance.
Recently it got some POD100 love as explained here.
At some point i removed the SB16 and LAN card from it to test something.
The system became unstable - not able to pass Quake1 test, sometimes hung in Doom, and other bad stuff.
Started reducing RAM, moving cache chips around, trying different video cards, tweaking BIOS settings.
No and no. I was like what the heck is going on - this one was always very stable. Did i damage it ?
At some point i put back the SB16. Instant stability.
What ?!
Adding more hardware usually means more volatile system - more things to go wrong.
Apparently not in this case. :)
Whatever.

That's just how it goes sometimes in this business, esp. with everything maxed out like this. L1 in WB with that VLB SCSI card is also a challenge.

I personally never liked those WD Caviar IDE drives. They were always failing on me back in the 90's. Moving to 68-pin Seagate SCSI drives was the cure.

Yes, is the drill.
Usually reduction is the way to improved stability, but this one was interesting.

Not big on WD drives either.
Had Caviars back in the day - rumble as thunder, run like a snail.
Later those 10k Raptors came and were big deal, but many of them had the lifespan of a gnat.

Adding an SB16 fixed problems? That's pretty weird. Maybe two pins in the ISA slot were making contact? I've experienced this myself a few times, and someone in this forum documented a case of this recently.

Cannot say if this change is on a logical or electrical level.
But is deterministic - SB16 off - unstable. On - stable. Didn't try with other cards to see if it is related to specific type of extension card, or ISA sloth/pins in general.

Asus ISA-386SIQ

A classic late model 386 motherboard, based on 486 class chipset SiS 461 - made me curious how this will work.
SiS 461 is pretty decent early 486 chipset - not too fast, not too slow - should be a good high-end fit for 386 system.

The board is basically jumper-less.
Supports up to 32Mb RAM and 256Kb SRAM.
Very stable.
Takes any components with ease. Maxed it out right away.
Used TI SXL2-50 CPU, Cyrix FasMath FPU (black top) and Cirrus Logic 5434 ISA VGA (Diamond) - pictures of them in previous posts.

BIOS is similar (same) as of 486 motherboards with this chipset.
It has very few options for adjusting timings.
Set to their optimum performance values.

Overclocks pretty well to 55MHz with FPU !

Looks like there were different production runs - some small differences in silkscreen printing, chips, etc.
Behavior is identical.

SpeedSys does not like SXL2 CPUs - text is all jumbled.
So going with just benchmarks today.

benchmark results

Performance is disappointing.
Didn't bother running 3D rendering tests. They take a while but won't reveal anything interesting. Picture is pretty clear already.
Tried swapping the BIOS with one from ASUS SV2 486 board - while it worked - perf stayed same.

This is not the first 386 motherboard from ASUS that passed through here and was lacking.
Apparently they were still sharpening their pencils at the time. 😀

---

UPDATE: It looks like the level 2 cache may be in always dirty state. User @jakethompson1 provided a work around in this post that may improve performance significantly. To be tested soon.

There is also a BIOS from Soyo SY-019J motherboard based on SiS 460 that has more robust control over SRAM/DRAM wait states than the standard ASUS one.
With it performance jumps between 20 to 30 percent in the different test cases. This is enough to move the board from being one of the slowest to intermediate performer.
Also, with the tighter wait states 55MHz FSB cannot be stable.
Didn't bother updating the consolidated performance charts - it is not worth it.
Maybe later ...

One last note, IBM BL3 processors don't work with this motherboard.

For quite some time i have been looking at who-is-who in the world of 386 mostly driven by personal curiosity with the end objective to identify existing, or preferably - build the fastest, fully stable 386 PC.
Motivation has been coming and going in waves, resulting in periods of varying activity.
Over time i accumulated enough data to be able to draw some conclusions.
Also, there is plenty of information around to fill any remaining gaps.
I stumble at various posts and articles, only to forget and rediscover them later.
Decided to put together a proper "expose" for future reference - when i forget the details again.

---

Ok, so here we go:

The variety of 386 grade components is large. It is impossible to inspect each and every one of them to draw 100% accurate conclusions.
So, let's define a search criteria as a starting point and then isolate the "culprits" of interest.

If we assume modern data storage like CompactFlach cards + adapters ...
Then assume decent quality RAM and SRAM - which are still easily obtainable ...
We end-up with 3 remaining factors that define the performance of a 386 system:
- CPU/FPU
- video card
- motherboard

CPUs
Consider systems that can handle 386DX CPU running at 40MHz, or higher frequency.
Also, include drop-in upgrades like 486DLC/SXL/SXL2/RapidCad. But also, IBM BL2/3 and next-gen CPUs through interposers.

FPUs
Technically Intel's RapidCad offers the best in this department, but the overall performance of the CPU falls behind SXL2 + Cyrix FasMath running natively at 50MHz.
Never had the chance to test Weitek Abacus 3167 and never saw any conclusive test results = unknown variable.
There are clock doubling variants (2x25 and 2x33) of IIT/ULSI/FasMath FPUs. In my experience they can operate reliably at 60MHz (possibly more), but performance tends to be worse than the FPU running natively at 50MHz. EDIT HERE: Based on Feipoa's numbers in the following posts - there is no clear line about this. Some tests benefit from the higher CPU/FPU frequency, others less so.

VIDEO CARDS
CL5434 seems to be the fastest ISA VGA - which is great fit for a clean 386 system.
For hybrid motherboards - Ark1000VL, S3 Trio64, WD90C33-ZZ, S3 96# deliver top perf in DOS and Win3x. 386 class CPU cannot saturate them, so any fast VLB card is fine really.

MOTHERBOARDS
This is the factor with the biggest variety of components that need proper filtering.
- ignore 386 SX systems - their performance is lacking compared to DX designs
- ignore early 386 motherboards - usually identified by wide format PCBs, low level of integration, small amount / no cache, no SIMM slots. There are few rare exceptions from that rule of course.
- ignore late 386 motherboards - the ones with very compact PCBs, highly integrated chipsets, usually 128Kb cache only, clock generators limited to 40MHz instead of crystal oscillators = no overclocking capabilities. Their niche was to fill low-budget markets given the fact that the world largely moved to 486 at the time.
- no FPU motherboards - goal is full featured, maxed-out, completely stable systems. No FPU = no-go.

Once apply the above search criteria, especially the "motherboards" one, the list of considered 386 components shrinks down to 20-30 items.
Spent the time to examine most of them and shared details about the more interesting ones in this thread.
Also, checked many derivatives/variations that didn't reveal anything new, didn't share details to avoid spam.

There are few bits of missing information that i was not able to abstain personally:
- VLSI's 82C331/332 chipset performance/stability and the based on it AMI Mark V (baby screamer) board.
- Had pretty good implementation of Forex FRX48C421 that fell silent before my 386 curiosity spiked, don't have captured data about it.
- Some Act and G2 386 chipsets that just don't inspire confidence by looking at the available implementations - early boards, low level of integration, etc. Maybe one day will check them for completeness.
- Cyrix 486DRx2-66 CPU.
- 486 interposers with 486 class CPUs.

---

Luckily some of that missing info has been covered by others:

In this post skv400 shares Doom scores with Aquarius MB-4D33/50 motherboard based on OPTi 495SX (or is it UMC UM82C491F chipset ?), Cyrix 486DRx2 66MHz and CL-GD5434 VLB 1MB VGA.
He achieved 16.39 fps in Doom.
This is the highest score in the entire giant 386 benchmarks thread.

386_junkie shares numbers here about pretty good looking Forex 386 board with 386DX CPU running at 50MHz and CL5434 ISA VGA.
Results:
9.57 fps in Doom
18.6 fps in Superscape
4 fps in PC Player Benchmark

JohnBruno posted here about his 386 ISA mobo sporting DRx2-33/66 CPU running stably at 40/80MHz with tight wait states.
He didn't reveal much perf metrics, but the two shared there are great.
Superscape - 37 fps
PC Player Benchmark - 9.2 fps

Predator99 published information about several other 386 boards here.

Feipoa talks here about Mark V Baby Screamer board with 486SXL2-66 CPU, but i am not sure what the outcome was. No perf numbers posted.
Then he is taking things to the next level with strapping 486DX4/5 CPUs on 386 board using interposers here.
If you don't mind the absence of a 386 class processor then this is the fastest system that includes a core 386 component (the motherboard).
So, in some ways i can end the post here. Mission accomplished. Done.

But to many this may be too big of a "cheat".
So let's keep going for completeness.

Feipoa again on 387 FPUs with benchmark results: thread

Kixs had a great start here with 386/486 OPTi VLB hybrid board and pretty much the full spectrum of 386 CPUs + upgrades, but never posted results.

Another interesting thread about 386 stuff, but no benchmarks/numbers shared.

---

If i consider everything mentioned above and compare the available perf numbers, this is what i arrive at (separated in 3 categories):

1. The 2 fastest clean 386 PCs seem to be the this one and this one.
All the magic of the first PC is in the DTK motherboard, based on a hybrid 386/486 Symphony Haydn chipset. The rest is more or less common stuff.
Notice, the clock-to-clock FPU perf is not the best out there but the additional 10MHz (over the base 40MHz ones) get things on the right track.
The second PC required magical components top to bottom. Not the fastest system on a clock-to-clock basis, but scales great.
Output performance and stability for both are excellent.

2. Only VLB motherboards with fast VLB video and EIDE cards can improve on some specific metrics/tests.
For example, IBM BL2/3 CPUs + Alaris Leopard/Cougar motherboards, showcased here and here.

3. It takes what Feipoa did with 486DX4/5 CPUs + interposers to climb higher.
So far the highest achieved performance i have seen on a 386 class motherboard.

3.1 Honorable mention for mpe's POD83 in a hybrid OPTi board here.
It uses the 486 hardware paths, but still great stuff.

---

So far this is what i see.
Did i miss anything ?

My performance numbers are somewhere on vogons for the Mark V Baby Screamer w/SXL2-66. I can re-run them if you wish. I know I have pages and pages of data tables for 386 benchmarks with different CPUs. The Baby Screamer is ISA-only. If you have a package with instructions on what to run, I can run it on my pre-assembled systems (no testbeds until my wife's standing desk add-on is gone).

In general, I try to limit my "ultimate 386" to motherboards which chipsets which are not also 486 chipsets. To some extent, these hybrid chipsets also feel like cheating. If one were to exclude those as well, then no VLB graphics cards either. So if your requirements are to have hybrid chipsets and VLB then The Mark V Baby Screamer w/SXL2-66 won't win that race.

Yes, I agree, the Am5x86-160 on a PGA-132 is too big of a cheat, but would have made a great upgrade for the time if the price was right.

I also have a Symphony 386 board waiting to be tested. It is E.jpg shown here (last board on first post): https://www.cpu-world.com/forum/viewtopic.php?t=33077

feipoa wrote on 2021-04-15, 00:59:

My performance numbers are somewhere on vogons for the Mark V Baby Screamer w/SXL2-66.

Exactly the reason i made the post above. 😀

feipoa wrote on 2021-04-15, 00:59:

I can re-run them if you wish. I know I have pages and pages of data tables for 386 benchmarks with different CPUs. The Baby Screamer is ISA-only. If you have a package with instructions on what to run, I can run it on my pre-assembled systems (no testbeds until my wife's standing desk add-on is gone).

That will be great. Thank you.
wolf3d, superscape, pc player benchmark, doom, wintune2 (this is more VGA dependent, but while at it ...)
Can send you the two 3D rendering apps ready to use.

feipoa wrote on 2021-04-15, 00:59:

In general, I try to limit my "ultimate 386" to motherboards which chipsets which are not also 486 chipsets. To some extent, these hybrid chipsets also feel like cheating. If one were to exclude those as well, then no VLB graphics cards either. So if your requirements are to have hybrid chipsets and VLB then The Mark V Baby Screamer w/SXL2-66 won't win that race.

I know what you mean. There are many ways to slice this.
I ended up relying on the PCB layout as a differentiator.
Clean 386 motherboard is something with 132-pin PGA CPU socket, ISA only, 30 pin SIMMs, dip28 sram chips, optional IDE on board.
No 486 sockets, no 72pin SIMMs, no VLB. I am less interested in these too, but they still partially fit the 386 bill, so added them here.

feipoa wrote on 2021-04-15, 00:59:

Yes, I agree, the Am5x86-160 on a PGA-132 is too big of a cheat, but would have made a great upgrade for the time if the price was right.

Agreed.

feipoa wrote on 2021-04-15, 00:59:

I also have a Symphony 386 board waiting to be tested. It is E.jpg shown here (last board on first post): https://www.cpu-world.com/forum/viewtopic.php?t=33077

So, 2 years already and the motherboard is still sitting on the reserves bench ?
😁

How is your Doom window set? Zoomed in all the way (no green border) and still having the status bar shown at the bottom?

How is your Wolf3D being run? Are you using wolf_286.exe timedemo?

I assume you are using PCPBench /VGAMODE ? How chart shows 640x480 and 320x200, but the test result is the same number.

3DBench original version, not v2? Your charts show it is 3dbench 1.0 and 3dbench 1.0c, but both results are the same. They should be different.

Are you running WinTune 2.0 in "full test mode" or "quick test mode"? If you go to properties, you'll see there is an option to select which.

I recall now the issue I had with 3D studio R3, I get a video error when I try to run it:

1Error Number 1400 : VESA initialization function 2 failed
2
3VESA failed to properly initialize the graphics mode. 
4
5You may be either trying to use a video mode not supported by this VESA BIOS or the VGA card does not have the required RAM to support the selected graphics mode. Try selecting a graphics mode with fewer colors or a lower resolution. You m ay have instailled the wrong VESA VBE (if required) for your VGA.
6
7Press any key to continue.

I press a key and it kicks me back into DOS and show the error again, but in fewer words:

1**** vdm_init mode 65 returned error 1400 ***
2Specified display too small
3Press any  key to continue

We are using the same GD-5434, right? How did you circumvent this problem?

I just finished installing Lightwave 3D v4, which is also available on archive.org. Even though I already had Win32 installed, the software wants MSVCRT20.DLL in C:\Windows\System. As you noted this test will take a few hours, I'll have to run it tomorrow.

Are you also able to post your 3DSR4 Chevy.3ds results? It is possible I won't be able to get 3dsR3 running.

I'll share the other results once these last two are complete.

Fixed the image in my previous post (some copy past stuff that was incorrect). Full refresh the page to get the update.

Doom - full screen with status bar.
Wolf3D - yes the timedemo. there is something a test.bat file in the version i have. I just run this.
PCBench - VGA mode only.
WinTunne2 - full test.

3DS
Yes, we are using the same CL5434 VGA.
Run 3ds vibcfg and then adjust settings. Render resolution should be 640x480. Set the rest to whatever you want.
If you are using clean install - high chance the vibrant driver is obsolete. Copy the one from the archive i provided.

LW3D
You can steal the DLL from the archive i provided.

3DS interface upon frame rendering completion is confusing.
These two screenshots bring the needed clarity.

(these are 3DS R3 render times)

Recently i obtained 4x4Mb 50ns 30pin SIMM modules from a forum member.
Put them to good use today with that DTK 386 system.
Was able to lower the DRAM wait-state from 2 to 1 cycles.
Here are some of the results:

Wolf3D: 61.1 fps
Doom: 18.03 fps <- !!!
Superscape: 34.4 fps
PC Player Benchmark (VGA): 8.9 fps

Wolf3D's status bar corruption comes from running cyrix -i1 -i2 command. It boosts performance with 0.2 fps, but at the expense of garbled pixels.
Without these flags picture is clear, but fps maxes out at 60.9.
They have no impact on any other tests/games/apps - both perf and image quality wise.

Doom crossed the 18 fps barrier on an 386 ISA system. Not bad.

I know that the PC is not fully stable in this configuration. It cannot complete some of the more advanced tests, but does pretty well in DOS gaming.
These vertical OC exercises are kind of pointless (to me) if there is no fully working computer at the other end, but thought it is worth sharing the "achievement".

With these SIMMs i can get things much more reliable at 55MHz, but wait states have to be increased, which results in lower performance than sharpened system at 50MHz.

I'll look at this again this evening. Keep in mind that my SXL2-66 w/AMI Mark V isn't shooting for the fastest PGA-132 system. Basically, the SXL2-66 works at 80 Mhz in every motherboard I tried EXCEPT the AMI Mark V Baby Screamer. For a system to really fly with the SXL2-66, it needs to be run at 80 MHz. But the novelty of the AMI Mark V Baby Screamer is enough for me to keep it as my "386 system". I have dozens of pages worth of benchmarks with the SXL2 at 80 Mhz and the BL3 at 100 Mhz, which I should really try to pull out some examples, otherwise such a data dump would overwhelm this thread.

In case you see my previous post as an attempt for horse-racing - not the point at all - just a coincidence while we are talking about 386 things here.
Couple of days ago these 50ns memory sticks arrived and it was time to give them a try.
They can contribute to some performance improvement in OC'ed configurations, but the system is not stable enough under such conditions (that take advantage of this RAM), so nothing to see there.

As for perf numbers and such from your MARK V system - you offered to share the common tests used around here + 2 rendering tests only.
That will be more than enough.

AMI Mark V Baby Screamer
32 MB 60 ns FPM
256KB L2 cache
TI 486SXL2-66 CPU running at 66 MHz
ULSI DX2-66 FPU running at 66 MHz
Cirrus Logic GD-5434 (actually a Photon Torpedo card running with the Diamond BIOS)
Adaptec 1542CP SCSI host controller card
SCSI2SD v6 with Sandisk Extreme Pro, V30, UHS-I, U3, A2 MicroSD; using firmware from 6.2.15

PCPBench /VGAMode
320x200x8bit
score = 9.8

3Dbench
score = 31.2

Doom (not using Phils back)
doom timedemo 3 with extra flags -nosound -nomouse
score = 4767 = 15.67 fps
Side note: The DRx2-66 score is only 13.4 fps. Does 8K vs. 1K make that much difference?

Wolfy3D
using wolf_286.exe
-all I did was copy the extra files from the Wolf_286 zip file into my official Wolf3D game folder. I did not copy ALL the files in wolf_286.zip because they appeared to have the same date and size as the original ones. I copied only the files from wolf_286 that I didn't have in Wolf3d.
-zoomed the view window all the way in (tiny green frame still)
score = 27.9 fps
comment: Our results are too different, so I suspect some settings are inconsistent

WinTune2
in full test mode
800x600x?-bit
the WinTune2 test result says it is running in 32-bit mode, yet I told the Diamond InControl Tools to run it at 16.7 million colours, or 24-bit. Hmmm..
score = 1198 Kpixel/s

3D Studio R3
render chevy at 640x480x8bit
score = 9:49, or 589 seconds

Lightwave 3D R4
render blade.lws, frame 50 only, at 640x480x8bit
score = 2hr 0min 2s, or 7202 seconds

For comparison, in checking my notes for MB's with hybrid UMC 481/482 with the same graphics card, same SRAM/DRAM quantities, etc.

1SXL2-50, ISA=8.3 MHz, DOOM = 5320 = 14.04 fps
2SXL2-50, ISA=12.5 MHz, DOOM = 5151 = 14.50 fps
3SXL2-55, ISA=9.2 MHz, DOOM = 4814 = 15.52 fps
4DRx2-66, ISA=11.1 MHz, DOOM = 4725 = 15.81 fps
5DRx2-66, ISA=8.3 MHz, DOOM = 4963 = 15.05 fps
6SXL2-66, ISA=8.3 MHz, DOOM = 4930 = 15.25 fps
7SXL2-80, ISA=10 MHz, DOOM = 3595 = 20.8 fps
8BL2-75, ISA=8.3 MHz, DOOM = 4200 = 17.78 fps
9BL2-75, ISA=12.5 MHz, DOOM = 4054 = 18.42 fps
10BL3-100, ISA=11.1 MHz, DOOM = 3152 = 23.70 fps

One more comparison, using a pure 386 chipset, SiS 310/330/330

1SXL2-50, ISA=8.3 MHz, DOOM = 5470 = 13.65 fps
2SXL2-50, ISA=10 MHz, DOOM = 5164 = 14.46 fps
3SXL2-55, ISA=9.2 MHz, DOOM = 4944 = 15.11 fps
4SXL2-55, ISA=11.1 MHz, DOOM = 4669 = 16.0 fps
5DRx2-66, ISA=11.1 MHz, DOOM = 4657 = 16.04 fps
6SXL2-80, ISA=10 MHz, DOOM = 3579 = 20.9 fps
7BL2-75, ISA=8.3 MHz, DOOM = 4690 = 15.93 fps
8BL2-75, ISA=10 MHz, DOOM = 4096 = 18.23 fps
9BL3-100, ISA=11.1 MHz, DOOM = 3068 = 24.34 fps

3D Studio just loves your 2x33MHz FPU.
Interestingly enough LightWave3D seems not to care much about it.

Thanks for the additional numbers.
They add definition to the blurry 386 hybrid business.

This is what i get with two Alaris boards + BL2-80 / BL3-100 CPUs:

Pretty much inline with your BL3-100MHz numbers.
I remember the exact Doom fps was a bit below 24.4. More like 24.3#, but closer to 24.4. Rounded it to the first decimal point.
While the Leobard BL2-80 system is fully stable, the Cougar BL-100 one is not.
Do you remember how stable things were during your OC runs ?

These tests were run many years back and I do not recall the specifics. I was attempting all perturbations of chipset, ISA freq, clkmul, 50 Mhz FSB, different FPUs, running Intel FPUs async, different oscillators, wait states, video cards, etc. There were too many combinations with thousands of datapoints tabulated. The hard part was finding a motherboard which would work with the BL3 upgrade modules in Windows. I finally found one, but the module covers the VLB slots :( This work was early preparation for a revised 486 benchmark comparison which would run from 286-12 to Pentium 100. I also have 286-to-486 upgrade modules to test and tweak.

My memory is such that the 3.6V SXL2 at 80 MHz had no issue in Windows 3.11 with whatever I ran on it. I as wanting to try 90-100 MHz but didn't get to it (oscillators in the mail took too long). I have one BL3 that is only stable at 75 MHz max, the other would run Windows at 110 MHz, but I didn't get around to stability testing. It had no problem in Doom or simple Windows tasks. It wasn't very interesting because on that particular module, I had to flush the cache at every I/O, rendering it rather pointless. I have another module that I know will do 80 MHz stable, but at the time, I couldn't test it in Windows because I hadn't found that magic motherboard yet. What we need is a proper IBM BL3 upgrade module made for IBM compatibles. I like the work RayeR is doing in this regard.

Hopefully somebody familiar with 4+ layer PCB design will design the custom interposer for these PGA168 variants of the SXL2-66. Custom interposer module for TI486SXL2-66 PGA168 to PGA132 - HELP!

My AMI Mark V Baby Screamer will work with the 3.6V SXL2 with a 40 MHz FSB, but the issue was that it wouldn't clock double at this frequency. The max frequency it would clock double at was 72/2. The 72 MHz oscillators I have get oddly hot, so I suspect the voltage is incorrect. I ran some tests at 70 MHz and 72 MHz, but unfortunately, I had that green bar around the doom screen so the numbers are artificially high. Some years later when I went to test it again at 70 MHz, the CPU wouldn't clock double. Hence it is left running at 66 MHz.

EDIT: Interesting how the SXL2-80 results benched higher than the BL3-80 results. The BL3 has twice the L1 cache.

Damn. I still want a PGA132 version of the SXL2-66, like the evergreen kit. Been looking for years. Never seen one for sale.

Seems they are rarer than a Cyrix 5x86-133/4x. Aren't there only 3 users on Vogons and another 3 on CPU-World who have these?

An ideal situation for me would be to put mine in a protective box for collection and run the SXL2-66 PGA168 variant on a PGA132 interposer.

Have you seen the QFP SXL2-66 3.6 V ICs NOS anywhere, kinda like how you found those SXLC2-66 ICs?