If you're going for pure performance, I'd say ditch the POD entirely and go with the P90, as it seems to perform much better.

Versatility wise, the POD box could theoretically scale better to games that require slower CPUs, since you could easily swap out the POD for a slower 486.

mr_bigmouth_502 wrote:

If you're going for pure performance, I'd say ditch the POD entirely and go with the P90, as it seems to perform much better.

Versatility wise, the POD box could theoretically scale better to games that require slower CPUs, since you could easily swap out the POD for a slower 486.

This setup is solely made for this project. It is not something I'll keep for daily basis. The Creative 3D Blaster VLB will be running on my 486 VLB board (with 1MB L2 cache) and there I will be forced to use the POD. I'm torn what will be the "best" way to compare NASCAR Racing against the VLB setup for the NV1 and Millennium as mentioned in my first post. This is were I'm asking for advise.

Let me list the pros and cons:

Using the 486 PCI board with the POD @ 100mhz :

Pros:

• Same CPU as on the VLB system for better direct comparison
• Same Socket 3 as with VLB, 486 "technology"

Cons:

• No L2 cache on PCI board make results not directly comparible
• NV1 and Matrox not primarily designed to run on Socket 3 486 PCI systems, but on Pentium's Socket 5 and 7.
• PCI cards have to run at 27mhz PCI bus at 40mhz FSB (can be solved by running tests at 83mhz (33mhz FSB)

As seen in the video clip the differences between the P90 and POD are noticeable in the accelerated Matrox version of NASCAR Racing (much more than in any other benchmark or game I've tried!).

I wonder, are there any Socket 3 boards out there that incorporate both VLB and PCI?

mr_bigmouth_502 wrote:

I wonder, are there any Socket 3 boards out there that incorporate both VLB and PCI?

http://www.phantom.sannata.ru/forum/index.php?t=2933
http://mail.lipsia.de/~enigma/neu/pics/mainboard_ga486im.jpg

Nice comparison video!

I'd go with the Pentium.

The P90 system also has more memory bandwith than the 486 board. That could make a huge difference.

iulianv wrote:

mr_bigmouth_502 wrote:

I wonder, are there any Socket 3 boards out there that incorporate both VLB and PCI?

http://www.phantom.sannata.ru/forum/index.php?t=2933
http://mail.lipsia.de/~enigma/neu/pics/mainboard_ga486im.jpg

I know these boards exist, but it is not an option to spend even more money on this at the current moment. There are also socket 5 boards with VLB, but they are rare. Would've been perfect for this test though....

Nice comparison video!

I'd go with the Pentium.

Thanks, more will come, that was just a short clip to show the dilemma!

I'm considering going for the Pentium myself, but the NV1 and Matrox will do better in the comparison as a result of this. I guess it boils down to how well do we want those two cards to compete against the VLB?

The P90 system also has more memory bandwith than the 486 board. That could make a huge difference.

That could be one of the reasons yes, but it didn't affect performance that much in Doom and Quake. Anyway, here are the Speedsys results on the P90 (POD 100mhz in parenthesis):

CPU: 67.37 (73.76)
Memory Bandwith: 184.6 MB/sec (121 MB/sec)
L1 cache: 177 MB/sec (195 MB/sec)
L2 cache: 90 MB/sec (NA)
Memory throughput: 62 MB/sec (45.4 MB/sec)

Just set up my 486 VLB system again for testing.

Hardware:
Asus VL/I-486SV2GX4 motherboard /w 1MB L2 cache
1x32MB FPM 60ns RAM
Creative Soundblaster AWE32 CT2760
VL-bus VGA Number Nine 9GXE64 S3 Vision 864 2MB

Here are the benchmark results:

I am actually quite impressed with beating Feipoa in Doom and Quake with my VLB system! I think it is pretty clear that I have to compare the VLB system with the 3D Blaster against the Pentium 90.

I also captured a new video showing my VLB system running in SOFTWARE mode against the Pentium 90. Beside the OBVIOUS VGA output quality differences between the S3 Vision card and the Matrox Millennium the systems are very similar in terms of performance. Only point were they really differ noticeably is at 0:40 in the video. Here there are alot of graphics being processed and I can imagine memory throughput and/or video bandwidth playing a factor.

NASCAR Racing in SOFTWARE mode on Pentium Overdrive (in VLB system) vs Pentium 90:
http://youtu.be/uqf9BXmpW74

Anything special need to be done to a POD to run it at 100Mhz (besides the 40mhz bus speed)?

NJRoadfan wrote:

Anything special need to be done to a POD to run it at 100Mhz (besides the 40mhz bus speed)?

Before I bought mine I asked the same thing. I was told only 1 out of 10 POD5V83's could run at 100mhz stable. I guess I was lucky..

It is as easy as setting the jumper to 40mhz FSB and voila 100mhz if everything works 😀

Have anyone tested performance differences between Socket 5 and Socket 7 boards with the same CPUs?

Are there any differences on running a later Pentium 166 downclocked to Pentium 90 compared to having a real Pentium 90?

Great comparison! I always get excited when people run their 486's to the max. A few comments:

It would be interesting to compare the NASCAR videos on the POD100 system on a board with 512 or 1024 KB of L2 cache and a full-speed 40 MHz PCI bus. I realise you don't want to spend any more money on this, so perhaps in time someone else will test this scenerio. I'd be surprised if the POD100 didn't meet P90's performance on a 430VX board. The conclusion of the Ultimate 486 Benchmark Comparison's Overall Performance chart was that the POD100 was marginally better than a P90. It looks like your VLB board with the POD100 seems to support this theory.

I am surprised you were able to run your POD100 in Quake1's timedemo. That was the only test I had to chill the POD prior to testing to get a successful completion. All other DOS and Windows benchmarks could be completed with out prior chilling. Only 1 in 3 of my POD's run well at 100 MHz, but it sounds like your POD runs even better than mine at 100 MHz.

Something else you might try is the AMD X5 at 180 MHz. This way you'd be running the PCI bus at 30 MHz and L2 cache is somewhat less important.

The Speedsys Memory Bandwith for your 486 VLB board seems erroneous.

Can you replace the fake/bad cache on your M921 board? Even if it is soldered, you can use a solder pump and remove the old cache.

vetz wrote:

I am actually quite impressed with beating Feipoa in Doom and Quake with my VLB system! I think it is pretty clear that I have to compare the VLB system with the 3D Blaster against the Pentium 90.

This is probably a consequence of the reduced PCI bus speed. For the 686 Benchmark Comparison, I beleive I ran the POD100 with a 40 MHz bus, but with a higher game resolution of 640x480. Could you let me know what your VLB system gets in Quake 1.06 at 640x480 (no sound, everything else at default)? I recorded 6.5 fps. What does your P90 get on the 430VX board at 640x480? I got an impressive 9.9 fps using a 430TX board and the P90.

vetz wrote:

Have anyone tested performance differences between Socket 5 and Socket 7 boards with the same CPUs?

Are there any differences on running a later Pentium 166 downclocked to Pentium 90 compared to having a real Pentium 90?

Never tried any of these but:

A) As long as the boards use the same chipset (not entirely sure if there ever was a chipset for both socket 5 and socket 7 though), RAM, L2 Cache and cards I don't see how performance would differ.

B) Again, performance should be the same. It's why Intel started locking the multipliers in CPUs. Stores would often buy something like a P166 and sell it as a P200 by changing the multiplier, just because it was stable enough (or not).

I was hoping you would reply feipoa 😀

feipoa wrote:

It would be interesting to compare the NASCAR videos on the POD100 system on a board with 512 or 1024 KB of L2 cache and a full-speed 40 MHz PCI bus. I realise you don't want to spend any more money on this, so perhaps in time someone else will test this scenerio. I'd be surprised if the POD100 didn't meet P90's performance on a 430VX board. The conclusion of the Ultimate 486 Benchmark Comparison's Overall Performance chart was that the POD100 was marginally better than a P90. It looks like your VLB board with the POD100 seems to support this theory.

Yeah, after I got the VLB board back up and running I agree with that. Would love to test with a fully functional PCI board with L2 cache, but if anyone else want to do it and post the results, please do 😀

feipoa wrote:

I am surprised you were able to run your POD100 in Quake1's timedemo. That was the only test I had to chill the POD prior to testing to get a successful completion. All other DOS and Windows benchmarks could be completed with out prior chilling. Only 1 in 3 of my POD's run well at 100 MHz, but it sounds like your POD runs even better than mine at 100 MHz.

While doing some testing to see if I could improve the results even more with BIOS settings tweaking Quake crashed. I presumed that was because I had put some setting to tight, but when you write this then it actually may have been the CPU. I will need to do some more prolonged stability tests to see if it really is an issue.

feipoa wrote:

Something else you might try is the AMD X5 at 180 MHz. This way you'd be running the PCI bus at 30 MHz and L2 cache is somewhat less important.

The PC Chips M921 only supports FSB up to 40mhz?! Is there some undocumented jumper settings to run it at 60mhz FSB?
There is nothing here: http://th2chips.freeservers.com/syl8884/
Though Amoretro says something about 66FSB here: http://www.amoretro.de/2012/07/pc-chips-m921- … otherboard.html

feipoa wrote:

The Speedsys Memory Bandwith for your 486 VLB board seems erroneous.

I agree it seems weird, but the memory throughput also seems very low. I also tested with cachechk and it reported 400ns read time on the RAM and 20MB/sec speed. That is low.. RAM problems? Or just error in benchmarking? Real life applications doesn't seem to indicate there is a problem.

feipoa wrote:

Can you replace the fake/bad cache on your M921 board? Even if it is soldered, you can use a solder pump and remove the old cache.

It is soldered on, but there are tracers and jumpers so I guess it is possible to remove it and solder on proper cache. I do not have the skill or the soldering tools to do this at the current moment though.

feipoa wrote:

This is probably a consequence of the reduced PCI bus speed. For the 686 Benchmark Comparison, I beleive I ran the POD100 with a 40 MHz bus, but with a higher game resolution of 640x480. Could you let me know what your VLB system gets in Quake 1.06 at 640x480 (no sound, everything else at default)? I recorded 6.5 fps. What does your P90 get on the 430VX board at 640x480? I got an impressive 9.9 fps using a 430TX board and the P90.

Yeah, I figured that the PCI bus speed was responsible for the lower results. I just tested the 486 VLB system in Quake 1.06. I got 9.3 FPS in 640x480 (!). I'll test the P90 later.

f2nbp wrote:

A) As long as the boards use the same chipset (not entirely sure if there ever was a chipset for both socket 5 and socket 7 though), RAM, L2 Cache and cards I don't see how performance would differ.

The chipsets will be different. I also believe the L2 cache was DIPs on socket 5 boards and soldered newer type of RAM on socket 7. I dunno if that have any practical implications. If I get hold of a Socket 5 board I'll use it for the video.

Refer to page 6 of the Biostar MB-8433UUD pdf manual found in the World's Fastest 486 link (see my signature). Here I list numerous unsupported FSB settings and the jumper patterns should be similar for other UMC 486 boards. I know my M919 worked with a 60 MHz FSB. From reference material found online, it seems J28 sets the FSB on the M921. There are 3 jumpers, which should hopefully lead you to all the unsupported FSB settings noted in my Biostar PDF manual. Of course, this also depends on which clock generator was used on your motherboard.

20 MB/s for RAM in Cachechk sounds really slow for a P100 on a socket 3. You should be getting around 42 MB/s. I have not really had any problems with cachechk and false RAM readings on a 486. Your VLB Speedsys marks are quite a bit lower than your M921 marks for memory throughput. Are the memory timings optimised in the BIOS?

For removing the fake cache on the M921, not a whole lot of skills are required. It is more like a test of endurance on the laborer's part. A $25 handheld solder sucker pump and a soldering iron are required. You melt the cache leads one at a time with the soldering iron; while the solder is molten, suck out the solder with your spring operated solder pump. The cache leads which go to ground usually require more time to head up before you can suck out the solder. I would recommend replacing the cache modules with DIP sockets from a dead 486 board so you can easily swap out the cache in the future. I also recommend practicing on some unwanted motherboards before doing any work on your wanted boards.

Yeah, I figured that the PCI bus speed was responsible for the lower results. I just tested the 486 VLB system in Quake 1.06. I got 9.3 FPS in 640x480 (!). I'll test the P90 later.

Interesting... I wonder if I did something wrong for the 686 benchmark comparison on these socket 3 CPUs. Can you use the Matrox Millennium and Matrox G200 to see what score you get with your M921 at 640x480? What could be choking my 486 at 640x480?

feipoa wrote:

Refer to page 6 of the Biostar MB-8433UUD pdf manual found in the World's Fastest 486 link (see my signature). Here I list numerous unsupported FSB settings and the jumper patterns should be similar for other UMC 486 boards. I know my M919 worked with a 60 MHz FSB. From reference material found online, it seems J28 sets the FSB on the M921. There are 3 jumpers, which should hopefully lead you to all the unsupported FSB settings noted in my Biostar PDF manual. Of course, this also depends on which clock generator was used on your motherboard.

Cheers 😀 J28 on my M921 is setup the same way as your Biostar, atleast physically. I'll test it out and report back with my AMD X5-133ADW

20 MB/s for RAM in Cachechk sounds really slow for a P100 on a socket 3. You should be getting around 42 MB/s. I have not really had any problems with cachechk and false RAM readings on a 486. Your VLB Speedsys marks are quite a bit lower than your M921 marks for memory throughput. Are the memory timings optimised in the BIOS?

Cachechk "write" values were the same as yours, it was just "read" values there were off. In the BIOS (I have the latest) there are no specific settings for DRAM WS for both read and write. I've set it to the lowest value possible, but maybe it only affects write values?

For removing the fake cache on the M921, not a whole lot of skills are required. It is more like a test of endurance on the laborer's part. A $25 handheld solder sucker pump and a soldering iron are required. You melt the cache leads one at a time with the soldering iron; while the solder is molten, suck out the solder with your spring operated solder pump. The cache leads which go to ground usually require more time to head up before you can suck out the solder. I would recommend replacing the cache modules with DIP sockets from a dead 486 board so you can easily swap out the cache in the future. I also recommend practicing on some unwanted motherboards before doing any work on your wanted boards.

Certainly a project for some other time 😀 Thanks for the info!

Can you use the Matrox Millennium and Matrox G200 to see what score you get with your M921 at 640x480? What could be choking my 486 at 640x480?

Ofc. Here are the results in Quake 1.06 at 640x480 with the M921 (POD 100mhz):

• Matrox Millennium: 6.4 FPS
  Matrox G200: 6.4 FPS

That is very similar to your result of 6.5 FPS. What could be the cause of choking? If you were running the tests at 40mhz PCI bus then I have no clue what could cause it. UMC Chipset "bug" perhaps?

Here are some more benchmarks of the Pentium 90 in Quake 1.06 at 640x480 as requested.
Matrox G200: 9.9 FPS (same as yours)
Matrox Millennium: 9.9 FPS

I also have a Socket 4 system with a Pentium 66.

Hardware:

• Pentium 66mhz (1x66)
  4x8MB FPM RAM
  Matrox G200 PCI
  Soundblaster 16 CT2230
  Intel OEM motherboard /w 256kb L2 cache

Here are the benchmark results of it:
Quake 320x200 no sound: 19.2 FPS
Quake 640x480 no sound: 7.3 FPS
Doom no sound: 42.4 FPS
3DBench 1.0: 71.4 1.0c: 70.6
PCPBench: 10.5
Speedsys:
CPU: 49.49
RAM Bandwidth: 100.7 MB/sec
VID Bandwidth: 46.6 MB/sec
L1: 124.8 MB/sec
L2: 60.46 MB/sec
Mem throughput: 39.1 MB/sec

The BIOS does not allow any memory tweaking, so it is running by default settings.

Perhaps some aspect of the VL-BUS allows for greater frame throughput at higher resolutions (640x480) compared to the PCI bus? I suppose one way to test for this is to use an M919, which has both PCI and VLB, however, I don't think there are many (any?) known good implementations of VLB and PCI on the same socket 3 board.

I don't think it is a UMC bug. I suspect a Matrox card on a SiS PCI board will yield very similar Quake 1 scores. Anyone want to test this?

Thanks for the P66 data. It was interesting to see that the benchmark results are about 15% less than a P66 socket 7 simulation.

Your guess is as good as mine when it comes to the 640x480 performance in Quake.

Anyway, I also tested the AMD X5 at 180mhz (3x60). Some interesting results!

Quake 320x200: 20 FPS
Quake 640x480: 6.2 FPS
Doom: 54 FPS
PCPBench: 9.9
3DBench 1.0: 99.9+ 1.0c: 99
Sysinfo: 389.9
Speedsys:
CPU: 67.44
Mem bandwidth: 182.31 MB/sec
L1: 171.27 MB/sec
L2: NA
Memory throughput: 68.86 MB/sec

I also tried 3x66mhz and 4x50mhz to run the CPU at 200mhz, but I could not get past POST. I even tried to increase the voltage to 4V, but with no luck. The results here show that if you get a AMD X5 to run stable at that speed it is quicker than the POD at 100mhz in a 486 PCI board in everything beside Quake 320x200.

The funny thing about running the AMD X5 at that speed is that SOFTWARE mode in NASCAR Racing were on par with Matrox accelerated mode on the Pentium 90! I'm pretty sure that it was as fast as on a Pentium 133 or 166! Matrox accelerated mode ran SLOWER than software with the AMD at 180mhz. I captured both and have footage of it if anyone is interested.

Anyway, I will not be using the AMD X5 at 180mhz for any comparison as it can't be run at that speed in my VLB board.

The AMD X5 at 180 MHz makes for a pretty nice system, even without cache. Of the 7 or so X5 chips I own, I cannot seem to get one operating stably at 180 MHz. What voltage and cooling solution are you using?

I did notice some mistakes in the U4BC and U6BC which I will need to correct.

For the U6BC, IBM 5x86, AMD X5, and Intel DX4 chip results had sound enabled at 640x480. fps went up only 0.2 or 0.3 fps. Likewise, for the U4BC, only the IBM 5x86C-133 previously had sound enabled. At 320x200, I now get 19.3 fps. I'll update these docs at some ponit. For comparison, a non-overclocked Cyrix 5x86-133/4x gets 18.2 fps at 320x200.

The funny thing about running the AMD X5 at that speed is that SOFTWARE mode in NASCAR Racing were on par with Matrox accelerated mode on the Pentium 90! I'm pretty sure that it was as fast as on a Pentium 133 or 166! Matrox accelerated mode ran SLOWER than software with the AMD at 180mhz.

Does the Matrox accelerated mode code require Pentium instructions to run properly?

At 69 MB/s memory throughput on the X5-180, you're still way faster than the L2 cache on a POD100. Maybe you want to think about setting up a system around the X5-180?