pshipkov wrote on 2023-03-07, 15:51:

@CoffeeOne […]
Show full quote

@CoffeeOne

None of that looks like a factor, but hey - we take the better perf. : )

@Anonymous Coward

Will give this another go.
My approach was not very structured in the first round.

Hi again,

I was lately "micro optimizing" my Asus SV2GX4 160 Ark Logic Setup.
Highest Doom Rate achieved is 1120 ticks => 66.7fps.
Also tried 60MHz external clock => Very bad idea, the board started to smell. The board still works, I am not sure if I damaged something.
I guess 60MHz vesa local bus is just not possible.

I tested again my Vesa Local Bus Chips & Technologies 64300 graphics card. And it now works (I don't know why it did not work when I first tested it. There is also a thread here on Vogons about it).
The doom bench got stuck with my default settings (everything on max). Then I changed "Latch Local Bus" to 3T and "Local Bus Ready" to Synchronize and the doom test worked (1134ticks => 65.9fps so not much slower than the ark logic 1000VL). Also Wolf 3D value is ok ~135 not too far off the the 153.3 with the Ark (Ark with 2T and Transparent).
What I like is that the picture quality is better under Windows 98 on my TFT in comparison to the Ark.
OK but now a question:
I did not find tests from you with the 64300. Did you have stability problems?
I put a second MB fast page mode in it, and also those 2 ram chips (256kx4) for the "XRAM".
I am not sure if it makes windows graphics really faster though.
The Wintune2 graphics test is faster with the CT at 800x600 16bit. 7500 versus ~6500 on the Ark.

EDIT:
CHIPS F64300 Vesa Local Bus - Video BIOS

Please share the tweaks you are applying to get ~67 fps in Doom. It is good to have them here.
Recently i tried to verify the 68 fps i achieved in the past, but couldn't - cannot remember anymore the tweaks used, so have to recreate the setup, which will take some time. Once there, will share the setup here.

Asus VLI has clock generator that can go above 50MHz, but the board is very unstable there.
But didn't notice any issues that can threaten the health of the board at these frequencies.
In a previous post i talked about 4x50 where things are mostly ok, but still far from full stability.
When it comes down to classic ISA/VLB boards and overclocking to 200MHz, check Chicony CH-471B - it is the real deal.

About C&T 64300.
I tested this card, but you are right that charts are not populated with the results, not sure how i missed that.
Will fix it, i still have the card and i believe the notes.
In general - it is nothing special. Ok in DOS and Windows. So, your findings are very much in line with what i have seen here.

pshipkov wrote on 2023-04-11, 15:38:

Please share the tweaks you are applying to get ~67 fps in Doom. It is good to have them here.
Recently i tried to verify the 68 fps i achieved in the past, but couldn't - cannot remember anymore the tweaks used, so have to recreate the setup, which will take some time. Once there, will share the setup here.
....

The most important thing is correct hardware setup, so L1 WB, VLB jumpers to 0 WS and <= 33MHz, ARK 1000 VL to 0 waitstates (or turbo mode which is available on my card).
Of course Bios with the "L2 Write Back Fix".
Bios setting all on max. So also "DRAM Write Burst" => enabled, brings a tiny improvment, same for "Slow Refresh" and "Hidden Refresh". ISA clock speed does not matter, so /3, /4 or /5 are all fine. /3 makes the network interface faster.

For Doom: Synchronize => 1133 ticks. Transparent => 1120 ticks. Also tiny improvement.

Very important: CPU must run in real mode (so no EMM386.EXE!). In protected mode everything is way slower. > 1200 ticks.

1st micro optimization: Get maximum of free dos memory (without using EMM386.EXE!), I can write an extra thread about that. BUT: Zero consequences for Doom and Wolf3D (so having 580000 bytes free or 590000 or 620000 is the same).
2nd micro optimization, but his one had an effect: Believe it or not, somehow diskspeed plays a tiny role. I know have a QD6580W controller with a speedup given by XTIDE Rom. XTIDE natively supportd QD6500 and 6580.
With my other UMC controller I had like ~4000 kBytes/s buffered read and linear read, now it is ~7000/6000. Then also smartdrv plays a role. For my setup (which is QD6580 controller + XTIDE ROM + Transcend compact flash card) smartdrv made it slower 😁 But all in all the effect of the diskspeed is like one fps only.

Having that written, I am quite sure, you know all that already.
The disk performance should be optimal for you already, because I remember you get something like 10000kBytes/sec or more.

How many ticks did you have with Doom? So 1120 is the absolute minimum (meaning highest frames per second) I can achieve.

And I was wrong with C&T 64300 doom value, for this I have 1177 ticks. Because I need to switch to Synchronize for stability.

EDIT: about Doom:
I tried now also the Trident TGUI 9440AGi. It runs at the same settings as the Ark. Very bad picture quality on my Eizo TFT though.
Doom: 1173 ticks, it beats the C&T 64300 in Doom. Amazing.

TGUI 9440AGI should have an integrated RAMDAC. Do you think the bad picture quality might be remedied by replacing or adding some filter components?

@CoffeeOne

These are all the right moves.
Same here, but there was another tweak, if i remember correctly for peak interactive graphics performance.
It was something with the jumper configuration. Something that is not according to documentation.
Now, i am worried that this sounds like a rumor, so don't bet $ on it just yet.
Will try to confirm over the weekend.

I don't remember the exact ticks that result in 68 fps in Doom (with the rounding), but it should be around 1100 or something like that.

---

I don't remember Trident AGI degraded picture quality catching my attention.
This card is real nice to me, but this can be the specific model. So hard to say if all other implementations are like that.

pshipkov wrote on 2023-04-13, 04:53:

@CoffeeOne […]
Show full quote

@CoffeeOne

These are all the right moves.
Same here, but there was another tweak, if i remember correctly for peak interactive graphics performance.
It was something with the jumper configuration. Something that is not according to documentation.
Now, i am worried that this sounds like a rumor, so don't bet $ on it just yet.
Will try to confirm over the weekend.

I don't remember the exact ticks that result in 68 fps in Doom (with the rounding), but it should be around 1100 or something like that.

---

I don't remember Trident AGI degraded picture quality catching my attention.
This card is real nice to me, but this can be the specific model. So hard to say if all other implementations are like that.

I had problems with the Am5x86, multiplier 4 was not working. I though it is the jumper settings. But it was the BIOS, you need BIOS 2.x for it.

So I asked for the jumper setting and you sent me this:
Re: 3 (+3 more) retro battle stations

That is all after manual except the first:
JP 16 should be 1-2, 4-5, but you sent me 1-2, 5-6.

Could be that it? I guess not.

Maybe, but not sure. I converted the vli pc to pod100 from am5x86 and the config i had is no more.
Since then i setup a second vli board for testing with am5x86, but perf is below the old peak numbers.
Will try to figure it out on the weekend.

pshipkov wrote on 2023-04-13, 06:39:

Maybe, but not sure. I converted the vli pc to pod100 from am5x86 and the config i had is no more.
Since then i setup a second vli board for testing with am5x86, but perf is below the old peak numbers.
Will try to figure it out on the weekend.

You have so many postings here on Vogons, it is difficult:
Re: 3 (+3 more) retro battle stations
What about these settings on the picture?
😁
EDIT:
Seems to be the same like here:
https://www.petershipkov.com/temp/retro_pc_im … gx4_rev_2.1.jpg
OK, so additional to the changed JP16:
You have here JP27 different (wrong), should be 2-3 for 1MB cache, on the picture it is 1-2.
And HW trap:
JP5, JP6 should be 1-2, 1-2 for Am5x86, but on the picture there is 2-3, 1-2 which means old Cyrix.
So three deviations: JP16, JP27 and JP5.

Notice that the first picture shows revision 2.1 of the board with 1Mb L2 cache size, so JP27 is in position 2-3, while the second picture shows revision 2.0 of the board with only 256Kb L2 cache size, so JP27 is in position 1-2.

About hardware trap - the most recent manual lists 5x86 as JP5 2-3 and JP6 1-2, unless i don't understand well the messy abbreviations of the CPUs and the 5x86 is not referred to Am5x86 but for Cyrix 5x86. Thinking about it now - that's probably it.
But also, i remember trying different configurations of these two jumpers without obvious difference.
Do you see a difference on your side ?

pshipkov wrote on 2023-04-14, 06:45:

Notice that the first picture shows revision 2.1 of the board with 1Mb L2 cache size, so JP27 is in position 2-3, while the seco […]
Show full quote

Notice that the first picture shows revision 2.1 of the board with 1Mb L2 cache size, so JP27 is in position 2-3, while the second picture shows revision 2.0 of the board with only 256Kb L2 cache size, so JP27 is in position 1-2.

About hardware trap - the most recent manual lists 5x86 as JP5 2-3 and JP6 1-2, unless i don't understand well the messy abbreviations of the CPUs and the 5x86 is not referred to Am5x86 but for Cyrix 5x86. Thinking about it now - that's probably it.
But also, i remember trying different configurations of these two jumpers without obvious difference.
Do you see a difference on your side ?

Hello,

I believe that all jumpers are identically between 2.0 and 2.1 (only the voltage selection is different at the 2.1: The 2.0 only has 3.45V and 3.6V, the 2.1 has 3.3, 3.45, 3.6, 4.0).

Why do you think JP27 should be on position 1.2?

About the hardware trap: I never used 2-3 and 1-2, because it is only for (non Intel S compatible) Cyrix. And for the non Intel S compatible AMD.
I don't have such CPUs.
Both Cyrix 5x86 and Am5x86 should use 1-2, 1-2. As far as I understood, it is about telling the mainboard which CPU type it is. which is important for working L1 write back cache.
Cyrix (old type) L1 WB can be controlled by Bios. Intel and AMD L1 write back cache cannot be controlled by Bios / software. Both Am5x86 and Cyrix 5x86 use the "new intel way".

Ok, but obviously the write back cache works also with the other setting. Maybe I should try it 😁

J27
because one of the pcitures shows 256Kb level 2 cache configuration.
check the first sentence in my previous post.
: )

Btw, i checked the HW trap and in reality jumpers are set properly to 1-2 position.
Not sure why they were off when snapped the photos back then.

pshipkov wrote on 2023-04-14, 23:26:

J27 because one of the pcitures shows 256Kb level 2 cache configuration. check the first sentence in my previous post. : ) […]
Show full quote

J27
because one of the pcitures shows 256Kb level 2 cache configuration.
check the first sentence in my previous post.
: )

Btw, i checked the HW trap and in reality jumpers are set properly to 1-2 position.
Not sure why they were off when snapped the photos back then.

Yes, right. Sorry, I overlooked that.

My latest DOS benchmark values with the VL/I-SV2GX4 Rev.2.0 with AMD 160MHz, Octec Ark Logic 1000VL 2MB.

1Quake low:		    18.3
2Doom max details:	1120 ticks
33D Bench 1.0c:		105.3
4Chris:				125.3
5Chris 640x480:		31.2
6PCP:				28.3
7PCP 640x480:		11.1
8Wolf 3D:			153.8

What's a bit weird, I have notes of a Quake value 18.4. Now it is always 18.3. But all other values are on top for me (I believe at least).
I played around also with 50*3, thinking that one or the other benchmark (graphics related) will be superior to 40MHz bus speed.
But: No chance! For example for the doom value: I don't even come close to the 1120 ticks, either > 1200 ticks or the system crashes during the doom bench run.
so the 150MHz is no more interesting for me.

LATE EDIT: Of course quake is NOT 640x480, when you choose this setting, the program falls back to 320 automatically.

CoffeeOne wrote on 2023-04-15, 14:12:

My latest DOS benchmark values with the VL/I-SV2GX4 Rev.2.0 with AMD 160MHz, Octec Ark Logic 1000VL 2MB. […]
Show full quote

My latest DOS benchmark values with the VL/I-SV2GX4 Rev.2.0 with AMD 160MHz, Octec Ark Logic 1000VL 2MB.

Copy code to clipboard

1Quake 640x480:		18.3
2Doom b:				1120 ticks
33D Bench 1.0c:		105.3
4Chris:				125.3
5Chris 640x480:		31.2
6PCP:				28.3
7PCP 640x480:		11.1
8Wolf 3D:			153.8

What's a bit weird, I have notes of a Quake value 18.4. Now it is always 18.3. But all other values are on top for me (I believe at least).
I played around also with 50*3, thinking that one or the other benchmark (graphics related) will be superior to 40MHz bus speed.
But: No chance! For example for the doom value: I don't even come close to the 1120 ticks, either > 1200 ticks or the system crashes during the doom bench run.
so the 150MHz is no more interesting for me.

You sure about that quake score? Sure it wasn't 320x240 🤣? My equivalent system is getting like 7fps in 640x480

maxtherabbit wrote on 2023-04-15, 17:33:

CoffeeOne wrote on 2023-04-15, 14:12:

My latest DOS benchmark values with the VL/I-SV2GX4 Rev.2.0 with AMD 160MHz, Octec Ark Logic 1000VL 2MB. […]
Show full quote

My latest DOS benchmark values with the VL/I-SV2GX4 Rev.2.0 with AMD 160MHz, Octec Ark Logic 1000VL 2MB.

Copy code to clipboard

1Quake 640x480:		18.3
2Doom b:				1120 ticks
33D Bench 1.0c:		105.3
4Chris:				125.3
5Chris 640x480:		31.2
6PCP:				28.3
7PCP 640x480:		11.1
8Wolf 3D:			153.8

What's a bit weird, I have notes of a Quake value 18.4. Now it is always 18.3. But all other values are on top for me (I believe at least).
I played around also with 50*3, thinking that one or the other benchmark (graphics related) will be superior to 40MHz bus speed.
But: No chance! For example for the doom value: I don't even come close to the 1120 ticks, either > 1200 ticks or the system crashes during the doom bench run.
so the 150MHz is no more interesting for me.

You sure about that quake score? Sure it wasn't 320x240 🤣? My equivalent system is getting like 7fps in 640x480

Hmm.

It is point e) from Phil's benchmark suite. Quake timedemo 640x480

EDIT: see pictures.
But I have no clue, is point e) only 320x240?

great numbers.
i thought i updated the charts with your top wolf3d score a while ago, but not seeing it in the images on the web server. will update tonight.

18.3 fps in quake 1 is for 320x240 resolution.

Max and others reported better perf with 3x50 than 4x40. i also see that one some mobos for some tests, but so far 4x40 is the better place overal.

Several years ago i examined Migitronics TK-82C491/386-4N-D02C, based on the highly integrated UMC UM82C491F chipset.
Back then the best 386 CPU i had was AMD 386DX-40. Also, even the board had wiring for crystal oscillators, it didn't cross my mind to remove the 80MHz limited clock generator and see where the overclocking limit is.
So, i have been curious since then how UM82C491F does with SXL2 / BL3 processors and if it does better than UM82C481/482 based boards.

Not far ago i stumbled upon ABit AB-AK3 rev.0.2 which is very similar to the Magitroncs one, but it died shortly after i started testing it with the more advanced CPUs.
Notes in the same thread linked above.

Finally obtained a working board that allowed me to answer these questions.

Magitronics TK-82C491/386-4N-D04

It is a slightly newer version of the 386-4N-D02C model - UM82C493F replaces bunch of individual ICs.
Very cool green sticker with broken English.
Minor battery leak that was taken care of immediately.
Removed the 80MHz limited clock generator, soldered crystal oscillator socket.
Removed the surface mounted AMD 386DX-40 CPU, installed PGA-132 socket.

Upper overclock limit is 45MHz. No lights past that.
Very stable at up to 45MHz, with one strange exception.
For some reason Windows often hangs with fast graphics cards such as STB Nitro or ET4000/W32i and installed driver with resolution above 640x480 + high/true color depths.
Especially with SXL2 and BL3 CPUs in clock-doubled mode.
Tried hard to figure out what is causing it, but without success.
I have to slow the system down to 33MHz FSB and 640x480 screen resolution for things to be stable.
I didn't have these problems with the 386-4N-D02C version of the board.

It is worth noting that in my brief testing ABit AB-AK3 rev.0.2 was able to go up to 55MHz, but anything above 40MHz was highly unstable.
So, it looks like integration makes quite a difference for these late compact assemblies, even they are based on the same chipset and have almost identical layout.

The AWARD BIOS is 486-like.
Lots of options.
It is faster than the Mr.BIOS microcodes for this chipset.

For the tests used trusted 128Kb 12ns rated ISSI level 2 cache chips and 50ns rated 16Mb RAM modules (4x4Mb). Funnily enough the board didn't like another set of trusted L2 cache chips from UMC.
Used STB Nitro video card with 2Mb RAM (Cirrus Logic GD-5434).

--- AMD 386DX at 45MHz, ISA bus at 11.25MHz

All BIOS settings on max, except:
AT BUS CLOCK = CLK/4 (best is /2)

Speedsys screen with FPU.

Very stable in DOS, unstable in Windows with driver with resolution above 640x480 + high/true color depths.
Intermediate performance.
Nothing more to add really.

--- TI 486SXL2 at 45MHz (1x45), ISA bus at 11.25MHz

Everything same as above.
Speedsys hangs.

--- IBM BL3 at 90MHz (2x45), ISA bus at 11.25MHz

Everything same as above.
Speedsys hangs.

---

benchmark results

Comparing the performance numbers with 386DX CPU between this board running at 45MHz and the previous version running at 40MHz - they are more or less the same.
Feels like performance regression.

Also, noticed that the board likes ULSI FPUs more than Cyrix FasMath.
Floating point intense tests are consistently slightly faster with ULSIs.

Not a bad board really.
Hustle free, but also - some issues that is hard to ignore.

Will call it a wrap with the 386 motherboard examinations.
I believe i touched on all (or at least most) major brands/assemblies/chipsets that could contribute to the peak-performance story.
Well with some small exceptions here and there, for example:
Forex 421 / 451, Contaq 591 / 592, Sarc, ETEQ Bengal, etc.

There are also early assemblies based on early chipsets from G2, CHIPS, OPTi, etc., that are too slow to make a difference.
Either skipped on them entirely, or didn't spam the thread with no-stories.

Going forward will be resorting to tactical 386 related updates only for the few remaining items of interest.

Thanks for doing this. I was always interested to see how some of these chipsets perform, but never had the resources to buy these items to test myself.
So I would say in general the hybrid 386/486 are not best performance options for the 386 (if ignoring VLB), but they are probably the most refined and hassle free.
I still stand by the opinion that 1992 was the year of peak 386. The best boards seem to be the older designs with added DLC support.

You are welcome. Effort was fueled by the same interest.

I share similar observations.
Early 386 hardware, before and around 1991 is too clunky.
Late designs - these very compact, budget motherboards, are too restricted. Cost cutting took its blood.
The period between late 1991 - early 1993 is where the best 386 designs emerged, or hit market. The very best of them are based on hybrid 386/486 chipsets, especially the ones with proper DLC support.

The 386 platform is one of the most fun areas in retro computing - hybrid 386/486 designs and DLCs make the whole thing quite convoluted and fascinating.
I used to be a platform purist, but Feipoa corrupted me. Once i accused him of that, but in turn he pointed the finger at you for corrupting him in the first place.
So this is how we go around here.