Chapter One: December 2018

• CPU: i486 DX/50
• Mainboard: QDI V4P895 256kB Cache (a very nice and stable board!)
• Graphics Card: STB Lightspeed VL 2MB (ET4000 W32p)
• RAM: 16MB 60ns FPM
• LAN: 3COM Etherlink III
• Sound 1: SB16 (CT1740) DSP 4.05 + Roland SCB55
• Sound 2: GUS Max
• SCSI Controller: Adaptec 1542CP ISA
• I/O Controller: Goldstar Prime 2 ISA
• HDD: IBM DCAS 4,3GB SCSI
• CD: Toshiba XM3501B SCSI extern
• Floppy: 3,5” + 5 ¼”
• Housing: Highscreen Colani Compact III

CPU: i486 DX/50
The goal was the challenge of a real 50MHz System with FSB=CPU. It should have a smell of high end which lead to an original i486DX 50MHz, no question.

Mainboard: QDI V4P895
The age of 486 was accompanied by several sorts and improvements of the bus systems (ISA, EISA, MCA, VLB and PCI). For me the VLB is a trademark of the 486, so my Socket 3 must have VLB. I don’t like VIP boards – from each village a cow, but nothing right. On the other hand, in my case, the 386 and 486 age was characterized by OPTI chipsets, so I want to have one system with an OPTI chipset. Good, that the OPTI895 is a fast and stable chipset with all necessary features. The QDI is a very good and compatible board with support of all known socket3 CPUs. And, very important, it has a good architecture, that supports more than three extra long cards and has a socket for coin cell (no Dallas or NiCd). I used for mine 12ns 32k8 cache chips.
Disadvantages are the missing support of 1024kB cache, no PS/2 mouse and the ugly AMI WinBIOS (but with all necessary features)

Graphics Card: STB Lightspeed 2MB (ET4000 W32p)
Nice and fast card for DOS. As a child I wanted this card, now I have one. Very compatible in DOS (except BattleChess 4000), but only poor refresh rates in Windows. This card has a 135MHz RAMDAC, but in Windows in 800x600x16bpp I got only 70Hz. Seems to be an issue with the driver. I tested several drivers, but found none with more than 70Hz support. And very important, this card supports 50MHz FSB without a WS.

RAM: 16MB with 60ns and Parity
16MB are more than enough for DOS and Windows 3.1. For 50MHz I need a fast (60ns) part and for high end one with parity, check! Tested with ctramtst.exe – superior stability.

LAN: 3COM Etherlink III
If an ISA LAN, the a 3COM. I don’t know why, but I ever want a 3COM, but never gave the money for that. Now 3COM is cheap and easy to get.

Sound 1: SB16 + Roland SCB-55
Sound is a very difficult thing. There is no perfect solution because of the several standards and some bugs in the implementation (e.g. HNB in SB16/AWE32). My sound solution should support an original OPL3 and an original GM. Because the games were programmed (mostly) with the references (Creative and Roland), you should have the reference if you want to have the experience the developer intended. For SBPro, SB16 and OPL support I take a SB16 (yes I know, no stereo in SBPro mode) without the HNB, means a DSP version 4.05, and for GM support I need a Roland solution. I don’t like the external “set-top-boxes” from Roland because it makes it more complicate to take my Retro out of the cabinet and play a game. Luckily the SB16 has a Wavetable header and there are solutions from Roland available for (SCB-7 and SCB-55). One big advantage of the CT1740 is the “plug and pray” free configuration. Set the jumpers to the needed positions and you need no TSR for initialization. On the other hand you have no problems with reshuffled resources if any PnP Tool makes botch. I was a fan of Jumper in this time, because plug and play caused many problems and bad surprises. And in the end I need no TSR which costs valuable kB of conventional RAM.

Sound 2: GUS max
The GUS was not really necessary but I always wanted to hear the Demo “Second Reality” with the original. And yes it sounds noticeable better than on SB. The GUS has a high coolness factor, but in the most games the sound of my SB16+Roland feels much better. Formal for the high end character I spend my GUS the upgrade to 1MB.

SCSI Controller: Adaptec 1542CP
I have many good experiences with SCSI. I started my SCSI career in 1994 with the CD-ROM (Toshiba XM3501B) and a Mustek scanner. If you follow the (simple) rules, it is very reliable with very few problems (e.g. no capacity limit by MB BIOS), and by the way, very fast (e.g. lesser CPU load). My high end system must get a SCSI system. Because of the 50MHz FSB a second VLB card is forbidden, so I took the latest ISA version. Adaptec has a good performance and currently a very good driver and BIOS support for their old components. Additional it gives me the elegant availability to add an external CD-ROM without violating my design-housing.

I/O Controller: Goldstar Prime2
I need one for the serial and parallel ports, had no preferences. But this one is good documented and reliable.

HDD: IBM DCAS 4,3GB
An old PC must have a spinning, clicking and rumbling Harddisk. In my Retro is no place for a Flash Disk. This special disk is my own, old disk from my earlier Socket 5 Pentium. With 4GB the size is sufficient to cover all my old stuff (ca. 1,7GB).

CD-ROM: Toshiba XM3501B
A very solid high-end CD-ROM with caddy from 1994. It reads the current AZO-CD-Rs from Verbatim. Especially the caddy is a very nice high end feature with a unique experience of this time. I own one in the end 1994 and scrapped it in the early 2000er as I switched from SCSI to SATA. This one I received via ebay. The external housing I “borrowed” from a Yamaha burner.

Floppy: 3 ½” + 5 ¼”
For my typical 486 I need both. Especially the reliability of the 5 ¼” disks was the reason why I kept this drive long time into the Pentium age. And I like the Floppy seeking while boot-up with both drives.

Housing: Highscreen Colani Compact III
This housing is from the German dealer Vobis with their trademark Highscreen. Especially this design, made by Luigi Colani, polarized the people. It has several functional disadvantages (function follows design), but it flashed me in my childhood and I was jealous of a friend who had one. Now I got one, uniformly yellowed in very good condition from ebay, very nice. The removable IDE-HDD case is used to store the keys for Keyboard and HDD.

Chapter Two: April 2019(changes only)

• CPU: i486DX/2 66 ODPR
• Graphics Card: Miro Crystal 20SD VL 2MB
• SCSI Controller: Adaptec 2840 VL

Apart from the fact, that the necessary wait states for cache and RAM @50MHz forced my board into transfer rates like 33MHz without wait states, the performance in Privateer wasn’t enough for my claims, especially in the end levels with much asteroids. So I decided to update to the typical 486, the DX2/66, with the side effect, that I can use my VLB SCSI controller in parallel to the graphics card.
In the mean time I got my good old Miro Crystal 20SD on ebay, as a substitute for the ET4000 W32p. Now I have 85Hz in Windows and this card supports 50MHz without wait states too.

Chapter Three: Storm the Summit - November 2019

While building this PC I tried out some other mainboards (with PCI and VLB) and CPUs (DX4, 5x86, POD) and got the feeling, there is more possible with my “typical” 486. The QDI supports max. 512kB cache, but only single banked, which I could not running stable @33MHz FSB. I also have the Biostar 1433/50-UIV, but it only supports 256kB cache and is significant slower than the QDI. In the end I found two fast and reliable PCI boards with 1024kB cache support, the Shuttle HOT433 with UMC8881 and the Chaintech 4SPM with SIS 496/497 and one with VLB, the ASUS SV2GX4.

In the meantime I got the idea to build up, based on a good old LION portable housing from my uncle, a solid 1992/93 386 PC. I found a similar board to my old one in 1993, a Shuttle HOT 307H, on ebay and made my first tests and benchmarks with it. Yes it is a fast 386, but it is too slow against any cheap 486. After I found out, that the dirty tag bit is not set correct, I decided to drop the 386 idea and make another authentic 486 (with deactivated L1 its nearly a 386) ISA system with OPTI chipset (495B) and 1993 focus into the portable PC housing and to max out my socket 3 system – great, two flies with one flap.

Back to my Colani Socket 3 desktop, I decided to stay with VLB because it is something unique for this period. PCI I have in all other of my Retros – boring. The fastest VLB board in my box with support of 1024kB L2 cache is the ASUS SV2GX4 Rev.2.1. BUT it has no correct dirty tag support with L2 in WB (L2 cache strategy to AUTO in BIOS setup sets always the L2 to WT). The SIS 471 supports a dirty tag directly via pin on the chipset. After several weeks of weigh whether I should risk this great, unique, running mainboard I did some benchmarks with my POD and tried to overclock it to 100MHz. Yes, it run stable, also in Quake, several hours. Ok, the dirty tag problem must be fixed!

The theory is simple, the tag cache needs an additional bit, which is controlled by the two additional signals (dirty tag and write enable) from chipset. Therefore I have to hook up an additional chip on top of the tag chip and connect all pins apart from the WE and all the I/O pins. The WE and one (doesn’t matter) of the I/O pins are then connected to the corresponding pins of the chipset (pin133 I/O / pin134 WE). This fix should run with the most other chipsets too (other pins). I was successful, now I have native dirty tag support independent of the BIOS implementation.

The VLB is very critical with 40MHz FSB and two Cards (VGA + SCSI), but I got stable function with SCSI controller in the VLB port narrow to CPU and the VGA card in the other. The WS and delays on VLB are all set to zero (means =<33MHz).

With L1 in WB I could not boot from SCSI HDD (@2840 and 1542), only with an IDE HDD (@Prime2 ISA). Seems to be a problem with DMA transfer and L1 WB. With IDE HDD the system runs stable with L1 WB. I read the datasheet from the SIS471 and the P24D and P24T are treated equally. So I tried some jumper configurations from the P24D, but with no positive effect for L1 WB behavior of my P24T.

For the Benchmarks and comparisons I switched to IDE. Later I will set the L1 to WT and switch back to my SCSI and live with the speed decrease and the knowledge what is practically possible.I know, the AMD 5x86 @160MHz and the Cyrix 5x86 @120MHz are comparable in performance to my configuration and are more “486” than my Pentium, but they are intended as budget solutions and would leave a bowl taste for me, more than the change from L1 WB to WT with the P24T.

The Final configuration of my Socket 3:

• CPU: P24T @100MHz (L1 WT)
• Mainboard: ASUS SV2GX4 1024kB L2 Cache (BIOS : 402.1)
• Graphics Card: Miro Crystal 20SD VLB 2MB Spea V7 Mirage P64/VL (S3 Trio64) 2MB
• RAM: 64MB FPM (2x 32 MB) 60ns w. Parity
• LAN: 3COM Etherlink III
• Sound 1: SB16 (CT1740) DSP 4.05 + Roland SCB-55
• Sound 2: GUS Max 1MB
• SCSI Controller: Adaptec 2842VL Adaptec 2842A
• I/O Controller: Goldstar Prime 2
• HDD: IBM DCAS 4,3GB SCSI
• CD: Toshiba XM3501B SCSI extern
• Floppy: 3 ½” + 5 ¼”
• Housing: Highscreen Colani Compact III

Chapter Four: Benchmarking

To save time, I used for the benchmarks 2x 8MB FPM PS/2 Simms (w/o Parity).

In the end I maxed out the BIOS settings (Chipset Features):

• DRAM Speed: Fastest
• DRAM Write WS: 0
• DRAM Write CAS: 2T (1T runs stable in DOS, ctramtst.exe certified a good RAM stability, used for benchmarks; for superior stability it must set to 2T and caused a minor <1% decrease of benchmark results)
• Hidden Refresh: enable (compatible without any problem and gives appr. +1% performance)
• Slow refresh: disable (slows down the access to Floppies, but accelerates the system a bit, for benchmarks used enable)
• L2 Cache: Write Back
• L1 Cache: Write Through (doesn’t matter, the jumper JP21 sets the strategy, WB mode used in Benchmarks)
• Cache Burst Read: 1T
• Cache Write Cycle: 2T
• Local Bus Ready: synchronized (transparent used for benchmarks, speeds up the SVGA modes a lot, but causes a reproducible crash in the end of the second sequence of “Second Reality” Demo, all other tested programs are not effected)

For the benchmarks in general i used the BIOS settings as follows with the exception of my final config.

As reference I did the tests with the P24D (DX/2 66) and the M1sc (Cyrix 5x86 100MHz, all stable registers enabled). The P24D supports L1 in WB too, but in this configuration the benchmark results are a little bit worse than in WT, strange. I also added the graphs with my earlier configurations with QDI (L1 WT, L2 256kB WB).
My final Configuration is the position 3: P24T 100MHz / L1 WT / L2 WB / VL synchr.

In addition to the Graphs I get 9,3fps in Quake @640x480 in my final configuration. With the maximum Config i got 10,4fps in Quake @640x480. Sorry, that i tested with V1.08.

Benchmark Setup:

• OS: MS-DOS 6.22
• Driver: HIMEM.SYS, Adaptec ASPI Manager, GUS driver, UniVBE5.1 for LFB support
• Doom 1.19, fullscreen with UI, -nosound, -timedemo demo3
• Quake 1.08, fullscreen with UI, -nosound, timedemo demo1

... and the mandatory SPEEDSYS screenshot ...

I am sure, some of you are interested in the performance of the POD @100MHz in other late 486 Mainboards with fast chipsets (SIS and UMC). To make it comparable I took only boards with 1024kB cache and L1 WB support, used a Miro Crystal 20SD PCI 2MB similar to the VLB version, and I used the same RAM (2x 8MB) and the same IDE HDD with the same SW. The only change was the usage of the onboard IDE controller instead of the Prime2 ISA. The PCI were overclocked to 40MHz.

• Shuttle HOT 433 – UMC 8881 (the lower results in Doom staying preserved also with other CPUs like M1sc and P24D and another S3 968 graphics card)
• Chaintech 4SPM – SIS 496/497 (PCPBench 640x400x8 tested without LFB, because it is massive slower with LFB – 9,9fps)

The boards got the newest, official available BIOS and I maxed out their configurations to the maximum stable limit (not documented by me).

At the end the comparison between the S3 Vision 864 and the ET4000 W32p – they are equal in DOS.

Final Words
The route is the goal. This project took in sum one year time and I made several experiences of my childhood history again – great. Now I have a system with appr. Pentium 90 Performance in a 486 platform to the price of multiple P90 systems – no problem. This system includes nearly all my childhood dreams and many nice impressions and gives me a good feeling. It’s my perfect Socket3 – for the moment.

Thanks for reading!

Great results. My POD100 systems is quite a bit slower than this (Doom 53fps, Quake 22.8fps).

That's with the pretty much same VGA (Hercules brand ET4000 w32p). The cache and motherboard features (transparent mode) seems to be making a difference.

I’ve often wondered about setting the pod to 1x multiplier and installing it to one of the rare undocumented 60mhz FSB 486 PCI boards, installing modern 10ns cache and seeing what happens

mpe wrote:

My POD100 systems is quite a bit slower than this (Doom 53fps, Quake 22.8fps).

That's with the pretty much same VGA (Hercules brand ET4000 w32p). The cache and motherboard features (transparent mode) seems to be making a difference.

Makes sense, a upgrade from 256kB to 1024kB gives appr. 5%, and apart from the difference between synchronized and transparent Quake 1.08 results appr. 5% better than 1.06.

Wow, this is really the perfect Socket3! I really like your build — very well thought out and tidy. I wish I had your skills in cable management 😀

Hope you get a lot of fun out of this machine.

While I understand why you wanted a true 50Mhz CPU after all these years DX2/66 VLB IS what a 486 is to many.
Problem is why limit ourselves to this when a 5x86 or POD are not difficult that to find either.

Even though I'm not a fan of the case design, the PC is a thing of beauty overall.
Saw the SB16 + Roland, thought nice, didn't notice the B till you started talking about the card and thought Very nice (Although I do like the display on mine)

And agree with the floopy seek noise at boot. Most my PC's have a gotek but couldn't do it on my childhood 486. After years of hearing it boot with a floppy it didn't sound right booting without it.

I actually love the choice of Colani case. Reminds me that Luigi Colani whi died two months back 🙁

Perhaps the case just needs removing yellow cast from the front panel.

jheronimus wrote:

Hope you get a lot of fun out of this machine.

Yes, planing to play Tie-Fighter SVGA in christmas holidays 😀 And thanks, but the Cable manaement looks better than it is ...

jheronimus wrote:

While I understand why you wanted a true 50Mhz CPU after all these years DX2/66 VLB IS what a 486 is to many.
Problem is why limit ourselves to this when a 5x86 or POD are not difficult that to find either.

This is a question like the "meaning of life". In my case any step was a epiphany of old memories and feelings. I get regularly the question from my friends: "why do you do that? there are emulators and a 10 year old smartphone ist faster ...?" - little minds! 😉
In the end i stay with a 486 DX/50 as an ISA system in addition to my Socket3 Pentium - I don't wanna miss it.

jheronimus wrote:

Even though I'm not a fan of the case design, the PC is a thing of beauty overall.
Saw the SB16 + Roland, thought nice, didn't notice the B till you started talking about the card and thought Very nice (Although I do like the display on mine)

Yes, this housing has a strong polarization to the people- you like it or you hate it. Me for myself in my childhood, i loved it with the first look, but i learned it hating as my friend wanted to upgrade it with a CD-ROM. Good example - it gives me the feeling from my childhood back. Sounds like midlife crisis 😉

mpe wrote:

Perhaps the case just needs removing yellow cast from the front panel.

The yellow looks more worse than it is. It is very evenly and is (for me) not not disturbing the look. And i don't know what happens with the Colani Signature while bleaching. I decided to keep it as it is.

Thanks all for the recognition!

Well Done, good job and thanks for the benchmarks!!

I like your Colani system 😀

There are not many Colani systems, i've seen the last few years here on vogons system specs.

Impressive work. Good job.

Recently somebody else posted about this motherboard.
It is one of my favorites too - one of the best clean VLB mobos out there.

I also noticed that it struggles with WB cache, so very interested in your hack. It will be great if you can expand on it further - detailed photos will help. How this will work for 512Kb of cache for example ?
Currently i run the board in WT mode, because in WB it is about 10% slower. Been bugging me out.
Didn't notice difference in performance regardless if i plug 256, 512 or 1024 Kb SRAM.
But i run the board with 32Mb of RAM only, so 256Kb cache covers it well.
I found that the POD running at 100MHz is stable for DOS games but fails at more involving tasks.

Thanks !

For 256kb and 512kB cache u use a 32k8 tag-cache in DIP28 housing, for 1024kB cache you use a tag cache in DIP32 housing. The dirty tag is an additional bit to the tag, so you can use the tag cache as a host for the dirty tag. Therefor yo use all signals and pins from the tag cache - except the pins for Write Enable (WE) and the I/O Pins. - The red marked Pins in the picture.

The easiest way is to use a similar chip for dirty tag like the tag chip and "push" it on top the tag, so that the pins overlap. Then or before cut the legs for WE and I/O of your dirty tag so that the legs dont touch the belonging legs of tag-cache. Then solder all other overlapping pins/legs.

The most difficult part is the connection to the chipset, because of the close distance of the pins. In the datasheet of the SIS471 you can see, that the pin for the alter bit (=dirty bit) is pin 133 and this for the alter pin write enable (WE) is pin 134.

Pin 133 is connected to a pull up resistor nearby and you can connect the wire to the terminal of the resistor. Pin 134 for WE is open and you have to connect your wire direct to the pin of the chipset which needs a fine soldering tip and a calm hand. Please consider to intercept the wires with glue or tape, so that no forces influences the soldering joints.

In the end shorten the wires and connect it to the corresponding (and open) pins of the dirty tag. Which of the 8 I/O pins you choose doesn't matter.

For comparision i did some benchmarks. These benchmarks are not with optimized BIOS settings so they are not completely compareble with this in my early posts.

In addition to the benchmarks i did some measurements with cachechk and speedsys

Cachechk (L1; L2; Memory) / Speedsys (L1; L2; Memory)
L2-Strategy: WT - Cachechk (115,2 MB/s; 50,8 MB/s; 36,2 MB/s) / Speedsys (162,01 MB/s; 44,35 MB/s; 37,15 MB/s)
L2-Strategy: WB (no dirty) - Cachechk (115,2 MB/s; 50,8 MB/s; 16,6 MB/s) / Speedsys (162,02 MB/s; 48,16 MB/s; 30,74 MB/s)
L2-Strategy: WB (with dirty) - Cachechk (115,2 MB/s; 50,8 MB/s; 36,9 MB/s) / Speedsys (162,57 MB/s; 48,71 MB/s; 37,41 MB/s)

The effect of missing dirty tag bit is as higher as smaller the cache is (its a fact of probability aof cache read misses and the following flushes), with the effect, that WT is often faster than WB without dirty. Finally i cannot understand why the developers of the mainboards didnt use the WB, espacially in later 486 chipsets, which provided to use a bit of the tag as the dirty (with the effect that the cacheable area halved).

Isn't the dirty bit actually fundamental to the operation of WB L2 cache? Without it, can you even call it writeback mode?

Anonymous Coward wrote:

Isn't the dirty bit actually fundamental to the operation of WB L2 cache?

Yes! And it has a relatively huge performance impact. This is why i cannot understnd, that many boards don't supoort dirty tag correct.

Anonymous Coward wrote:

Without it, can you even call it writeback mode?

You can! But there is no performance advantage against WT. On the contrary WT is often faster.

It is not actually.

As I understand the benefit of WB without tag is that the CPU doesn't have to synchronously wait for write operation as in write-through mode. Just when the cache line is about to be evicted there is a forced write (even if not needed).

Depending on implementation and circumstances this might be a good trade-off or not (especially considering you often get double max cached RAM size when disabling the dirty tag).