My adventure into 386 upgrades started in the mid-90's when I purchased a second-hand 386 board from a friend. It had an AMD 386DX-40. For a long time, I always thought this was the fastest CPU for the ole 386 motherboard. Much later, I later realised that there was a Cyrix 486DLC which also ran at 40 MHz but contained 1 KB of on-CPU cache.

More time passes and I was informed about the Texas Instruments 486SXL, which is nearly identical to the 486DLC but has an increadible 8 KB of L1 cache. Naturally, I wanted to take the machine to the max, so I upgraded again.

More time passes and I was able to source a clock doubled version of the 486SXL-40, called the 486SXL2-50. I paid the $25 and upgraded yet again. I didn't really like how it ran with a 25 MHz FSB compared to the SXL-40, with a 40 MHz FSB. Obviously, the memory and L2 cache run slower with a 25 MHz FSB. However, recent testing has revealed that this is of little consequence compared to the faster CPU speed of 50 MHz. Some boards really shine with the SXL2-50, while others show only marginal improvement above the SXL-40. I am still working on a large 386 chipset benchmark comparison with all these high-end CPUs, so stay tuned.

Originally, I was running the SXL2-50 with a Cyrix FasMath 83D87 at only 25 MHz. This bothered me. Unfortunately, only Intel i387DX FPUs can run asynchronously. None-the-less, I sourced the Intel i387DX and ran it at 40 MHz. To my dismay, FPU benchmarks ran slower with the i387DX at 40 MHz compared to the Cyrix FasMath at 25 MHz. There exist clock-doubled FPUs from IIT and ULSI. [side note: IIT's are slightly faster than ULSI clock-for-clock] I sourced an IIT x2 and ran it with the SXL2-50. Luckily, the FPU results were faster than the single Cyrix FasMath.

Wanting more out of my 5V SXL2, I was able to establish that running it at 55.25 MHz is stable. Unfortunately, the IIT x2 couldn't cut it at 55 MHz. So it was back to the Cyrix FasMath. Luckily, the SXL2-55 combined with the Cyrix FasMath at 27.6 MHz (55.25 / 2) is equal to the IIT x2 at 50 MHz, so no loss.

There was also a low voltage version of the TI486SXL2 which ran at 3.6 V. It came in PGA168 and QFP144 formats. Shown above is the QFP144 variant placed onto a PGA132 upgrade module, which also contains a voltage regulator to drop the motherboard's 5 V down to 3.6 V. It is rated for 66 MHz (clock-doubled), but it also runs well at 2x40, or 80 MHz. The QFP144 to PGA132 upgrade module is difficult to find, however the PGA168 CPUs are readily available even today. For this reason, it would be nice if there was an upgrade module which converted PGA168 to PGA132, specific for the 486SXL2-66 PGA168, e.g. Custom interposer module for TI486SXL2-66 PGA168 to PGA132 - HELP! Unfortunately, the SXL2-66 would not work in the Transcomputer module.

Some 386 boards have built-in support for enabling and supporting cache coherency between the DLC/SXL's onboard L1 cache, while some require software or hardware work-arounds, e.g. adding an external NAND FLUSH circuit, using the BARB register to invalidate the cache at every hold, or merely enabling via software L1 cache and opening up the cacheable region. Thus far, all my motherboards work with the L1 cache of these CPUs using software to enable it.

There was also a Cyrix 486DRx2 which rated speeds of 50 MHz and 66 MHz. These chips supposedly have the cache coherency circuit built into the unit. Unfortunately, they only have 1 KB of L1 cache. Is a Cyrix 486DRx2-66 with 1 KB of cache faster than a Texas Instruments 486SXL2-50 with 8 KB of cache? They are pretty darn close, but it depends on the motherboard. On my VIA 481/495-based board, the SXL-50 is a good bit faster than the DRx2-66, however on my UMC 481/482 board, the SXL2 needs to be run at 55 MHz to catch up to the DRx2-66. On my SiS 310/320/330 board, the DRx2-66 is faster than even the SXL2-55. More on this chipset vs. CPU comparison in the months to come.

Recently, I acquired an IBM Blue Lightning BL3, which is similar to the DLC/SXL/DRx, but added 16 KB of L1 cache in write-back mode (as opposed to write-through mode on the DLC/SXL/DRx). The IBM BL3 also added dozens of fancy register settings which CTCHIP34 is able to modify. Like the DLC/SXL/DRx chips, the BL3 still contains a 386 core with 486 instructions and requires an external FPU. There is a driver utility from Evergreen Technologies and one from Kingtston which can setup these registers for you at boot. For what register settings I use with which chip, refer to this thread Register settings for various CPUs

The BL3 contains a variable multiplier of 1x, 2x, and 3x while the DRx2 and SXL2 contain a 1x and 2x multiplier option. The Buffalo BL3 module would not run at 2x40, but ran at 3x33 MHz. The BL3 module shown in the middle would run at 2x40, depending on which motherboard it was placed in. It would also POST at 120 MHz, however it was not stable in Windows 3.11. I have sourced a 70, 72, and 74.5 MHz crystal oscillator so that I can see if it is stable at 105, 108, and 112 MHz, respectively.

Lastly, there was the Intel RapidCAD, RAPIDCAD-1 SZ624 and RAPIDCAD-2 SZ625. I believe the RapidCAD was essentially a 486DX-33 CPU without any L1 cache. The lack of L1 cache cripples performance significantly.

Was this the end of the road for 386-to-486 upgrades paths? Apparently not. I've been sitting on these Transcomputer modules for a few years now and haven't tested them until recently. They didn't come with jumper headers for CPU selection, so I added them. They send 5 V to the CPU, so if you are testing these, test a 5 V CPU. I originally tested the Cyrix DX2-66 on my UMC 481/482-based 386 motherboard using the smaller 486HPi module with jumpers set for a DX2. The system powered up. It could enter the BIOS to display the screen summary, but I could not navitage the pages. This motherboard has a "MR BIOS" brand BIOS in it and thinks the CPU is a Cyrix 486DLC. I beleive this confusion is why I cannot navigate the BIOS pages. Not important as I had already set the CMOS settings accordingly. To my surprise, the 5V Cyrix DX2-66 ran beautifully and the L1 cache was already enabled by the BIOS. Next I tried a Cyrix 486DX4-100, however I did not have an additional interposer module setup to run other chips.

I took a VRM interposer module which came with a Gainbery IBM 5x86c-100HF and modified it to be a variable voltage VRM. I figured this would come in handy since there are Cyrix 486 CPUs which run at 3.3, 3.45, 3.6, 3.7, and 4 V.

The Cyrix DX4-100 also ran beautifully with L1 enabled. When I installed an Intel DX2-66, the CPU ran fine, but the L1 cache was disabled as determined by cachechk and chkcpu. I was also able to browse the BIOS with the Intel DX2 installed, but there was no option to enable the L1 cache. Does anyone know of L1 cache enabling software for Intel/Cyrix/AMD DX2/4 CPUs? All I could think of at the time was SetMul.exe. I ran setmul.exe L1E. The program spit out a pretty nasty error, but according the CHKCPU and CACHECHK, the L1 cache is enabled. I confirmed L1 was working by running varous benchmarks.

I figured it was about time to get serious and pulled out my Evergreen AMD Am5x86-133, which is already on an VRM module.

Here is my setup. I am running an Am5x86-160 with 32 MB RAM, 256KB cache, Diamond SPeedStar 64 (Cirrus Logic CL-GD5434 w/2MB), Adaptec 1520B SCSI, 3Com Etherlink III 3C509B, and a generic Sound Blaster Vibra16.

MR BIOS actually measures the CPU's speed correctly on the startup screen

CHKCPU also records the correct information

Cachechk indicates a step at 16 KB, confirming that the L1 cache is working.

Interestingly, Speedsys seems to think that L1 cache is not enabled. Can someone explain this and explain the strange spike on the graph? I found in the past that Speedsys can have problems with odd CPU configurations and I tend to trust cachechk more.

Below are the results from Quake (timedemo 1), DOOM (timedemo 3), PCPBench (vga mode), and 3DBench.

How do the results of the 386-to-X5-160 compare to results of an AMD X5-160 on a PCI 486 motherboard? For simplicity, I will use "Quake 386" to refer to the results of my UMC 481/482-based motherboard running the AMD Am5x86 (x5) at 160 MHz and "Quake 486" to refer to the results on my UMC 8881/8886-based PCI motherboard.

Quake 386 = 9.7 fps
Quake 486 = 17.3 fps

DOOM 386 = 32.4 fps
DOOM 486 = 64 fps

PCPBENCH 386 = 6.1
PCPBENCH 486 = 9.6
[using VESA Modus 100 640x480 8bpp LFB]

3DBENCH 386 = 55.5
3DBENCH 486 = 94.1

Ideally, I would also want to run the PCI 486 motherboard with the same ISA graphics card (CL-GD5434) and compare again.

However, what I can estimate now is what performance the Transcomputer system has to that of the PCI system with a given CPU. From these results, the Transcomputer Am5x86-160 has the gaming performance of a PCI-based 486 system with an Intel DX2-66-WB. Actually, it falls somewhere between an Intel DX2-66 and a DX4-100. The ISA graphics is really holding it back. For non-graphics benchmarks, I suspect the results are a lot closer to that of an Am5x86-133, but I've yet to test this. It would be really interesting to see a test done with one of those VLB-based 386 motherboards.

Another good test to run is Windows 3.11. If you have having cache coherency issues, it will often show up while booting Windows or from within Windows. I did not have any problems. Here are the results.

With the SXL2, if the L1 cache isn't playing nice with the system's cache controller, usually you will have DMA issues with sound cards and crashing (this, by the way, is fixed by setting the proper non-cacheable regions). So with the Am5x86-160 installed, I opened Winplay3 and played a mp3 file in stereo, 44 KHz, and 128 mbps without any down sampling. I ran it for 2 minutes without a single skip.

Wanting to test my luck, I opened up Corell Word Perfect 6.1 and typed away. There weren't any skips. Word Perfect openned as if there wasn't a CPU hogging mp3 being played.

Surely if I did a search with Internet Explorer 3 there would be skipping...

nope! And IE opened and ran Google Search as if I was running a multi-tasking OS. I was shocked.

This concludes my testing for the time being. In the image of the system setup, you may notice a white wire emerging from the motherboard. This is connected to the ISA pin MEMW#. The other end, I was thinking needed to connect to J1 on the Transcomputer module for proper cache coherency, however connecting it or not did not seem to make a diference, and as such, I left it disconnected from the module. It was mentioned to me that there was an ISA card sold as an option with the Transcomputer which connected to J1 on the Transcomputer module. I haven't seen this ISA card or know how it is configured, but I beleive it is for enabling proper cache coherency on particular motherboards. I was just guessing that it connects to MEMW#. If anyone has one of these ISA cards, could you share a photo?

I haven't tried all my 386 motherboards with the Transcomputer modules, but I did try one other board, a SiS Rabbit-based motherboard (SiS 310/320/330). This board would not show any signs of life with the module in the CPU socket.

I also tested the other Transcomputer module - IBM 486PX. This is a much larger module and rises above the first ISA slot. Intel DX2 or DX or SX chips would not work properly with this module. It POSTed but couldn't run anything. I suspect it has something to due with the L1 cache being disabled by default. The Cyrix DX2/4 ran just fine with it though. The performance results between the two Transcomputer modules were the same. DOOM scored 3077 in 486PX and 3071 with 486HPi. The Cyrix DX4 seemed to run fine at 120 MHz and runs quite cool in comparison to the other CPUs.

I wish I had bought more of these 486HPi modules. I had no idea it would work so well and was so rare.

I will be testing my other 386 boards as well. I expected more of an L1 cache problem with these modules. I suspect the ease may have been from the fact that the UMC 481/482 also served as 486 chipset. I am looking forward to testing with CHIPS and VLSI boards. If those boards have issues, but at least boot up, I noticed that the Cyrix DX2/DX4 chips still have a BARB register, as indicated by CTCHIP34. The BARB register can be invoked to enable L1 cache flusing everytime the CPU enters a hold state. This "software" method is sometimes used with Cyrix DLC and TI SXL CPUs to get the L1 cache working.

Before anyone asks, I should point out that I have not had the courage to install a Cx5x86 into the Transcomputer module. I'm not wanting to push my luck at this stage.

Reserved for future updates (Summer 2017).

Is the Drx2-66 on the SiS faster than the SXL-55 on the VIA?

The VIA board didn't show any benefit of the SXL2-55 over the SXL2-50 for some reason. I cannot explain it with certainty. All other boards did show benefit of the SXL2-55 though.

The VIA SXL2-50 scores 4773 (15.6 fps) in DOOM timedemo 3 using a 6.25 MHz ISA bus

The SiS DRx2-66 scores 4978 (15.0 fps) in DOOM timedemo 3 using an 8.3 MHz ISA bus
The SiS DRx2-66 scores 4657 (16.0 fps) in DOOM timedemo 3 using an 11.1 MHz ISA bus

The VIA board has no adjust-ability for the ISA bus speed, so 6.25 MHz is what you get.

I'm not sure if you are comfortable running your SCSI card at 11.1 MHz. If you are, then I suppose the SiS DRx2-66 is marginally faster than the VIA SXL2-50. Clock-for-Busclock, the VIA board is faster. Its even faster with 6.25 MHz vs. 8.3 MHz on the SiS.

I'm not really interested in re-obtaining a DRx2 33/66 for a few extra points in DOOM. Prices for the 66MHz model are bat shit crazy.I'll just make due with what I have on hand, that being an SXL-40, a 486HPi and a possibly not working blue lightning upgrade.
I was pretty impressed with the compatibility of the 486HPi. Although I only tested briefly just to make sure it was working, my system easily booted with all the CPUs I threw at it including a Cyrix 5x86-100. I think the am5x86-160 is the way to go though, especially if you can get your hands on a QFP module like what you have.

You are OK with 11.1 MHz on the ISA bus? 11.1 MHz is borderline for me. The major downfall of the SIS Rabbit is the inability to use the Adaptec 1540CP DMA SCSI controller with DLC/SXL CPUs. Even the DRx2 CPU failed, which supposedly has the cache coherency circuit built-in. It would still be good to add the NAND FLUSH circuit to the SiS Rabbit board and see if that helps the issue any. On that board, it must be added to DMAHLDA on the chipset. I recall using HLDA on the CPU pin alone caused some minor issue.

It was interesting, though, that using the BARB method (to invalidate the L1 cache) on the other motherboards (CHIPS 351, VLSI 330, VIA 481, UMC 481) made the Adaptec 1540C work just fine with the SXL/DLC chips. Special note: UMC w/DLC chips did not need BARB enabled for 1540C to function properly.

Do you know if the DRx2-66's built-in cache coherency circuit is the same as that included on the TI486SXL2 PGA-168 chip? If it is, then according to Appendix D3 of the TI486SXL databook, the NAND L1 cache FLUSH circuit is still required.

Did you try your SXL-40 at 2x25 MHz? I want to test mine, but it is sitting in my cased VLSI system. The CPU sits under the drive bays and is quite involved to take it out for a simple test like this.

Do you have benchmarks results when using the Cyrix 5x86-120? Maybe I should test that out too. I won't be putting the Cyrix 5x86-133/4x in there though.

Not okay with more than 10MHz ISA. I usually fall back to 8 just to be safe.

No idea about the cache coherency circuit in the TI486SXL2 PGA 168.

I did try my SXL-40 at 2x25, and it seemed to work. I didn't test it for extended periods of time though.

No benchmarks for the Cyrix. I tested at 100MHz, not 120.

I must say I truly enjoyed reading through your journey. I had no idea of all the upgrade options available for the 386, but it sounds like a lot of people were in the same boat and just moved on to the 486s. Props to your for the courage to test some of these out, not knowing if they'd work or not. At some point, you should post benchmarks for each step up, from stock 386-40 on. That would be neat to see, time permitting.

Thanks for taking the time to document this.

Impressive! That's what I was looking for to upgrade my 386 but it's an ULC based mobo with the 40Mhz sticker on it. Could I use these upgrade modules too on it?

That's pretty sweet, to be sure! I wonder how well say, OS/2 or WinNT runs on it?

Also, there were 386 boards with VLB? That's kinda cool, never seen one. I do have a funny 386 board that can also take a 486 (both sockets), but I don't even know if it works.

In any case, thanks for doing the research and heavy lifting here!

This guy reports on a 386 with VLB graphics.
Re: Post your 386 Speedsys results here

Now that's a skookum choocher, by 386 standards. Very cool. Some rather uncommon hardware was made and sold in the ROK, I wonder if those were ever sold in NA or Europe? Would be sweet to get hold of one (or a VLB Pentium, but those were more common and less stable, I'd bet.)

Thanks for your effort, feipoa, very instresting read.

jade_angel wrote:

Now that's a skookum choocher, by 386 standards. Very cool. Some rather uncommon hardware was made and sold in the ROK, I wonder if those were ever sold in NA or Europe? Would be sweet to get hold of one (or a VLB Pentium, but those were more common and less stable, I'd bet.)

They still show up on eBay from time to time. There was recently one with an ALi chipset which sold on eBay. It seems that they really need a clock doubled CPU to benefit from VLB graphics.

I would guess that 386 with VLB is quite a bit easier to find than a Pentium with VLB. 386/486 hybrid boards with VLB were relatively common.