VLB was certainly not designed for 40mhz or 50mhz FSB. VLB was one of the reasons DX40 and DX50 never really caught on and why the DX2-66 became legendary.

From the Redhill CPU guide:

Actually, VESA had a lot to do with the market failure of the DX-50. It was hard to make an ISA system run reliably at 50MHz, and almost impossible with VESA. PCI wasn't out then, and when it did come out it was dreadfully unreliable for the first year or so. By the time of the Pentium 75, 50MHz and even 66MHz bus speed was easily achievable: in the time of the 486DX-50 though, 40MHz was tricky stuff and 50MHz boards were a nightmare.

My Asus VL/I-486SV2GX4 (rev 2.0) is based on the SIS 85C471/407 chipset. It seems there are some kind of error when doing memory readings using the POD as both sysinfo and cachechk is giving strange readings. Here are cachechk:

It must have been a mistake by me with not setting the Cache Write Cycle to 2T. I changed it and I had no stability issues. PCPBench increased from 11.2 to 11.3 and Doom went from 55.4 to 56.7. Other tests remained at the same results. I've also updated my BIOS picture above with a screencap (and not one from my phone).
Link to BIOS: http://www.elhvb.com/mobokive/archive/asus/48 … x4/sv2g0402.zip

I've noticed the memory read speed to be really slow like this when the dirty bit is not set appropriately in write-back mode. Try setting your L2 cache to write-through mode and run cachechk again. I suspect this will give a proper speed reading. Are you able to pass MemTest86 4.00 with L2 in write-back mode? If not, what about in write-through mode? On a SiS 496/497 board I tested recently, 1024K had these really slow readings in cachechk in WB mode.

I check out the BIOS but there are no hidden BIOS chipset options. I was looking for a setting for the dirty bit, 7+1 or 8+0. When I have run into slow cachechk responses like this, the 7+1 setting seemed to correct for it.

MemTest86 4.00 with L2 in WB mode gave 0 errors. It reported a memory speed of 68 MB/sec, which I suspect is more close to reality.

On another topic the 430FX board I got hold of (Iwill P54Ti) was DOA. So to get this test done I've bought another 1995 430FX board (on the way in the mail), Freetech 586F52:

Should provide the most fair comparison for this small "project" which Gona also stated in an earlier post.

From what I've read COAST modules on 430FX boards wasn't available until Q1 1996.

I changed L2 cache to AUTO and the results in games jumped up considerably:

Quake 320x200: 26.1 FPS (+1)
Quake 640x480: 9.8 FPS (+0.5)
Doom: 57.9 FPS (+1.2)

3DBench 1.0c score fell to 96
PCPBench went down to 11.2 again

Speedsys memory throughput now shows: 45 MB/sec (a tad above the 486 benchmark comparison). The Memory Bandwidth still shows 245 MB/sec

Cachechk scores now normal (compared to 486 Benchmark scores):

With these scores I'm starting to doubt that the Pentium 100 (Pentium 90 is now too slow unless I compare against a 430VX/HX system!) in a Socket 7 430FX board is going to beat this monster of a VLB system.....

vetz wrote:

I changed L2 cache to AUTO and the results in games jumped up considerably...Cachechk scores now normal (compared to 486 Benchmark scores):

Auto probably set the correct dirty bit (7+1), which you otherwise do not have manual control over. If the system has it set to 8+0 for WB cache, I too get slow memory read performance.

feipoa wrote:

vetz wrote:

I changed L2 cache to AUTO and the results in games jumped up considerably...Cachechk scores now normal (compared to 486 Benchmark scores):

Auto probably set the correct dirty bit (7+1), which you otherwise do not have manual control over. If the system has it set to 8+0 for WB cache, I too get slow memory read performance.

Yeah, was thinking the same myself.

Anyway, the findings in this thread indicate huge performance differences on the early Pentium chipsets. It would be really nice to have a full comparison on all of them. It is apparent that you can't say a Pentium 100 is a Pentium 100 no matter what.

For the Iwill P54Ti you received, there is a red dot on the BIOS chip, which I assume means it is bad. Try using an EEPROM programmer and reflash the BIOS. I have accidentlly ruined EEPROM BIOSes in the past by forgetting to switch the jumper from the 12V EPROM setting to the 5V EEPROM setting. However, there are some EEPROM BIOSes from Intel which were also 12V. Anyway, after replacing the BIOs with a reprogrammed piece, the board worked fine.

We really do need a 430FX chipset board to complete off this comparison. There was a remarkable difference between the P100 on the ALI vs. that of the 430VX.

What I find particularly interesting is the Quake 1 demo1 at 640x480 and how you achieved 9.8 fps, whereas my board only went up to 6.5 fps, but your same board with no L2 cache achieved 6.4 fps. A factor of the VL-bus perhaps? I always got higher benchmarks with VLB cards compared to PCI ones on a 486.

Which socket 5 boards had a UMC chipset?

The picture I posted is of the Freetech 586F52 board (Intel 430FX). I haven't received it yet, but I sincerely hope that red sticker doesn't mean that the chip is bad... It could be a QC sticker! The Iwill P54Ti board I can barely find any information about online. No BIOS or manual available and since I have no EEPROM programmer available trying to getting a new generic BIOS to see if that is the problem. At the moment it does not even boot and whatever I try does not cause any reaction from the speaker (booting with no CPU or memory should cause the speaker to beep, but it is completely silent with the screen being black). Anyway, I only paid 5 dollars for it with local pickup and I can always use it to practice some soldering skills...

I guess it is partly the VL-Bus at 40mhz causing the increased performance. Though that is not the whole picture. Even the Socket 4 430LX with the Pentium 66 had better score in 640x480 than the POD100mhz in the 486 UMC board. There also seems to be some kind of performance dip in Quake in higher resolutions on 486 PCI UMC boards which we have theorized about earlier in this thread.

Here is a list over early Socket 5 chipsets that I've found (some chipsets were avaible as Socket 4 (SIS and Opti) and 7 (430FX) as well):
Intel 430NX, Intel 430FX, SIS ( SIS 85C50X ), UMC (UM8891AF), VLSI 590P, Opti (82C556/7) and ofc the ALI chipset already tested.

My second Freetech 430FX board also turned out DOA... Damn my luck 🙁 I've posted a wanted ad on Amibay incase anyone wants to help out.

Until I get a working 430FX board this project is postponed.

I tested my 430HX and 430NX boards with Pentium 90 and Pentium 100. Updated chart:

Goes to show that the 430NX chipset is really a dog when it comes to performance. Look at that difference in Quake and Doom compared to the 430VX/HX. It is also interesting to see how well the Socket 4 system keeps up with the Pentium 90 and 100 on the other two early chipsets.

430FX board acquired. With a Pentium 100 and aggressive memory timings with 60ns RAM I'm about 1 FPS below the VLB system. My guess is that when upgrading from onboard DIP-type asynchronous SRAM cache to pipeline Burst SRAM that will match the systems almost perfectly. I'll report back when I find my COAST sticks 😀

Coast stick found and systems tested. I also threw my TX and MVP3 board into the benchmark. Click the link for full view:
https://dl.dropboxusercontent.com/u/57506833/ … enchmark_v3.PNG

Notice the big difference on ASYNC cache and Pipeline cache on 430FX! SDRAM and EDO doesn't give much difference. MVP3 also doesn't perform very well compared to the Intel boards. 430HX wins, but this might be because my 430HX board is of better quality than my TX and the BIOS offers better settings.

For the testing I'll be going with 430FX at 100mhz with async cache. It is a slightly slower than the VLB system at some tests, and faster at other. The most important is that the PCI bus remains at 33 mhz and it is the same mhz as the VLB system.

This concludes this thread for my part. I might create a new thread with the benchmark results incase someone wants to help out with other Socket 5 or 7 chipsets. A small preview of what to come:

As requested, here the results on the ASUS SP97-XV

Thanks for the benchmarks Mau1wurf, but I'm afraid they cannot be used.

I've tested the S3 cards on my 430FX platform and they give quite different (lower) results than the benchmark "standardcard" (set by feipoa) Matrox G200 PCI which I used. Also Doom is missing one increase in framesize.

This also opens up another problem in this quest of mine to find an equal system to my POD 100mhz VLB. Comparing the VLB version of S3 Vision 864 with a Matrox PCI card is not easy as there are so many other variables. The results vary greatly on the same system just changing the graphics card. I think I need to get hold of the PCI version of the same card to give it any fair chance. All tests must then be run again. If I do the full retests I will do the whole "feipoa package" of benchmarks (ultimate 486 comparison) with the Matrox G200 (and the basic tests with the S3 Vision864 PCI).

Oh I always use the default setting as per the Doom benchmark website 😀

I think you will always run into a ton of variables with these kind of projects.

What one can do is sample enough data and then use statistics to find some meaning. In the end of they day I think it makes bugger all of a difference. The results are all so close together the way I look at it.

Mau1wurf1977 wrote:

Oh I always use the default setting as per the Doom benchmark website 😀

I think you will always run into a ton of variables with these kind of projects.

What one can do is sample enough data and then use statistics to find some meaning. In the end of they day I think it makes bugger all of a difference. The results are all so close together the way I look at it.

That is the default setting?? Christ... I thought I read some place that Feipoa used full screen (with UI), but now I can't find that info. Feipoa: Can you confirm screensize on your Doom testing?

Sample enough data is the key here. I thought I had it figured down, but then I realised I won't be testing software vs software mode, but accelerated mode on two different cards. To get an more equal comparison the rest of the system components (CPU, memory, cache) need to perform almost equal in practical applications. I did a quick test with SVGA software mode on a 430FX system with a S3 Virge card and pipeline cache. Here the POD@100mhz system were a bit quicker over a 30 seconds test, but when I switched to the Matrox card it was very hard to tell them apart in side-by-side comparisons captures. Yes, I'm going for this kind of accuracy, with it being hard to tell them apart with the naked eye on side-by-side shots. The problem is that other benchmarks for CPU/system are very theoretical, like Whetstone and CPUmarks and whatnot. Here the results are abit similar in some benchmarks, but way off in others (like PCConfig and GraphicsMark). If this comes from incompatibilities I do not know.

So far that is similar:
Landmark 2.0:
POD 100: 550
P90 430FX PIPELINE: 517

CPUmark32:
POD 100: 180
P90 430FX PIPELINE: 203

Justin Benchmark:
POD 100: 2,54s
P90 430FX PIPELINE: 2,7s

Way off:
PCConfig v9.33
POD 100: 110
P90 430FX PIPELINE: 160

Graphicsmark (WinBench 96)
POD 100: 13,5 (with S3 Vision 864 VLB)
P90 430FX PIPELINE: 19,7 (with S3 Virge 325 PCI)
Results here are probably due to Virge being a better Windows accelerator. My guess is that any Matrox card with stomp on these scores.

http://www.complang.tuwien.ac.at/misc/doombench.html

Yes the standard setting of the 1.9 shareware version ins one border + HUD.

Mau1wurf1977 wrote:

http://www.complang.tuwien.ac.at/misc/doombench.html

Yes the standard setting of the 1.9 shareware version ins one border + HUD.

Some results on that list is too good to be true.

Also not using memory manangers like EMM386 and no mouse increases results. You had this loaded right?

I never use memory managers for benchmarking 😀 Not sure about the mouse either...

Keyboard + mouse is standard in the DOOM setup AFAIK.

I don't think EMM386 or HIMEM.SYS have any effect on Quake, PCPBENCH and 3DBENCH, but I'll test.

FEIPOA: Please answer regarding memory managers and mouse settings as well.

The Doom results in this thread can atleast not be compared to other benchmarks.....

vetz wrote:

I don't think EMM386 or HIMEM.SYS have any effect on Quake, PCPBENCH and 3DBENCH, but I'll test.

Without EMM386 & HIMEM.SYS on my POD 83@100mhz VLB system (loaded in parenthesis)
DOOM realticks: 1092 - 68,4FPS (1133 - 65,9FPS)
Quake FPS: 26,2 FPS (26,1FPS)
3DBench 1.0c: 96,4 (96,1)
PCPBench: 11,4 (13,9)

As you can see most scores had little difference beside Doom and PCPBench (VESA Modus 100 (640x400 8bpp LFB)). With PCPBench I tweaked some settings in BIOS to increase the score from what is in the previous post results table (full AUTO cache settings). It is strange that the score increases with memory managers loaded in PCPBench than without, which is completely contradictory to the other programs.

I think I need a W32p VLB card to further increase that DOOM score. Still pretty impressive 😀