FesterBlatz wrote:

Since I didn't see any decent/inexpensive Trio64V2+ boards on eBay here in the states I just snagged a STB Nitro Virge GX instead.

http://www.ebay.com/itm/231481352228

That's the same card I bought, same price too. Very nice. The Millennium will probably be a hair faster, but the GX is so much nicer with DOS games, compatibility-wise.

Modbin is a sinch. Don't forget to save the image. That step is not immediately obvious. Sometimes the modifications to the CMOS setup don't get saved. I think I had to save after each page of modification.

I'm just pointing out my experience with PCI 486 boards and the type of FLASH they use. You cannot really gage what you have by it. Sounds like you are going to check anyway, which is great.

I have often wondered about those Cyrix configuration register settings in MODBIN, however I have been unable to make a reliable map between what MODBIN shows and what are in the actual cyrix registers. The Cyrix DX4-100 also had settable registers, which were different from the Cyrix 5x86, so it seems tricky. Finding a proper mapping would allow us not to use the driver software to set the registers, which I suppose could be nice for using obscure operating systems.

Nice price for a GX. I have one with a "DELL STB NITRO 1.3" sticker over the VGABIOS. I'm not sure what DELL did to the BIOS. It came with 4 MB soldered on with "83 MHz" written on the RAM. I haven't bothered to try other VGA BIOSes because the Dell one seems fine. No black level bug anyway.

Ive got an unidentified ALI5119A motherboard an unfortunately appears to have fake cache 😠 .

Speedsys doesn't detect it either does cachechk, Ive disabled the external cache in the bios but the secondary boot screen still says "256" in the L2 cache section, the cache is soldered to the motherboard but I thought it was real as the benchmark results still seem pretty good?

Cyrix 5x86-120 with enhancements

Quake Timedemo1 - 15.8fps
Pcpbench 640x480 - 10.1fps
Pcpbench 320x200 - 22.3fps

P!ssed off!

Looks like you found a PC chips board.

Maybe you can do the same thing as I did here : PChips m915i looking for a bios which doesn't fake the cache

It's a long process (especially when you salvage sockets on another mobo) but for me it works so I think it was worth it

dirkmirk: Sounds like PC Chips strikes again! Does it have double-banked L2 cache? If so, I think it might be worth a desolder job.

Oh I forgot, before you do anything, make sure that the fake cache chips are wired up correctly because if they don't, this mod will be completely useless. On my board, I used a voltmeter on the cache legs and I saw that they were in fact powered.

To identify your board, search the POST string that is at the bottom of the screen when the computer is counting the RAM on google.

Also, are you really sure that the chipset is realy an ALI one ? Pcchips often put renamed chipsets on their boards. Oddly, on mine they removed the name (UMC chipset) and they put a sticker on it ... With the correct name written on it
duh 🤣

So over the weekend I did some benchmarks of the SMT Terminator with an Am5x86 at 160MHz (4x40MHz) using both EDO RAM and FPM.

Here's the Speedsys result with EDO RAM installed:

And here's the FPM results:

I also ran cachechk but I forgot to copy the txt files so I'll add them to this post later today.

I'll run the Quake and Doom timedemos once my S3 ViRGE shows up, which may even be today.

I also spent a lot of time using this board with one of my IBM 5x86c CPUs at 120MHz with EDO DRAM, and even with the lower clock it outperforms the AMD 5x86. In this case I did run a Quake timedemo, and with the Matrox Millennium installed I got 16.7 FPS.

I did some investigation regarding the CPU voltage selection jumpers since they don't function as described in the only manual I could find for a slightly different version of my board. As I suspected, the voltage regulator uses a TL431 followed by a small NPN that drives a TIP129 darlington. The jumpers select which Vout "R2" is selected. It turns out that as manufactured, the voltages are pretty far off from what is expected. The 3.3v setting results in about 3.1 volts, and the 3.45v setting results in about 3.5v. Calculating expected Vout using the installed resistor values matches my measurements, so I think this was an intentional compromise since the 4v setting puts both the 3.3v and 3.45v "R2" resistors in parallel to provide the 3rd voltage selection. All the testing above was done using the 3.45v selection which ran both the AMD and IBM 5x86 CPU's at 3.5v. At 3.1v the system was completely unstable, which isn't unexpected.

It is interesting that the L1/L2/RAM throughputs with your ALI w/X5-160 aren't any better than that with a UMC 8881 board w/FPM. e.g.

ALI
L1: 140 MB/s
L2: 56 MB/s
RAM: 44 MB/s

UMC
L1: 141 MB/s
L2: 59 MB/s
RAM: 45 MB/s

So I wonder if your motherboard is interjecting some wait states with FPM then removing the wait states when EDO is installed.

Cyrix 5x86-120 and AMD Am5x86-160 are usually pretty close in performance. What are your results with Am5x86-160?

Sounds like they really cheaped out on the board components for the VRM. To make a compromise for 3.1 V (which is useless) just to save money on an extra resistor? I usually like to remove the 4V resistor to make variable output VRMs on board. This way, you keep the 3.45 V setting intact.

I had a little time yesterday to run some more tests with the ALi chipset using my new STB Nitro S3 ViRGE (feipoa: if you want a dump of the STB video BIOS for your Dell version let me know). I will say this, the STB Nitro picture quality on an LCD display is superb, perhaps rivaling even the mighty old Matrox Millennium I bought brand new in 1996. The Doom, Quake, and PC Player benchmark results are identical to the Matrox (and NVidia TNT) too.

For some reason I was unable to get my Am5x86-133 to run reliably at 160MHz, even though last week it did so without a hitch. I'm not exactly sure what the reason is, but I suspect it may be CPU voltage related since believe it or not when overclocked it's MORE stable at 3.1v than it is at 3.5 or 3.6. And the chip was more than adequately cooled. I have a Kingston upgrade Am5x86 with the built in LDO and SQFP CPU package, I may retry with that.

So these are only the results for an IBM 5x86-100 running 40MHz x 3 at 3.5V with no signs of instability except when using EDO with the fastest timings. EDO seems to require dropping DRAM READ from "FASTEST" to "FAST" in the BIOS otherwise CTL-ALT-DEL reboots will often hang up the system and corrupt the CMOS BIOS settings. I left the timings at their fastest settings for the purpose of FPM:EDO comparison.

Results below. In a nutshell, it seems like using either EDO or FPM will shift the memory benchmark results around but gives no practical difference in performance.

IBM 5x86c-120MHz FPM S3 ViRGE

Speedsys Reslts
CPU: 67.59
Memory Bandwidth: 149.2
L1: 162.05
L2: 58.45
Main: 42.83

CacheChk Results
L1: 246.7
L2: 93.5
Main: 65.2

PlayerBench:
22.8

DOOM:
Gametics: 2134
Realtics: 1390

QUAKE:
16.7 FPS

IBM 5x86c-120MHz EDO S3 ViRGE

Speedsys Reslts
CPU: 67.58
Memory Bandwidth: 132.7
L1: 170.41
L2: 66.76
Main: 49.37

CacheChk Results
L1: 246.7
L2: 93.5
Main: 59.5

PlayerBench:
22.8

DOOM:
Gametics: 2134
Realtics: 1394

QUAKE:
16.8 FPS

FesterBlatz, do you have the Nitro v1.3 BIOS, or 1.1? Somebody saved the 1.1 BIOS for me, but I'd like v1.3 if you have it. If you bought what is in that EBay listing, it has 1.3 on a sticker, that's what I'm after.

Your story of decreased reliably when overclocking per different day is familiar; you might want to swap out your Am5x86 for another one. It may still be CMOS timing related, although, if your Cyrix 5x86-120 is stable, then it is probably not due to CMOS timings. Usually the Cyrix 5x86-120 is harder to get stable with fast timings than is the Am5x86-160.

I too had issues with the ALi board and fastest timings. But I really need to retest this with only 256K cache. The more cache you have, the less likely the fastest timings will work.

What speed is your RAM? Try 60 ns.

Is it only the Cyrix 5x86-120 which has issues with EDO and fastest? Does the CYrix 5x86-100 have issues with the fastest EDO setting?

I'm really surprised the Quake and DOOM results were practically identical. An 8 MB/s increase in cache speed should really have showed up in the benchmarks. Something is up. Could you also run cachechk -d -w -t4

This will check the RAM write times as well.

feipoa wrote:

FesterBlatz, do you have the Nitro v1.3 BIOS, or 1.1? Somebody saved the 1.1 BIOS for me, but I'd like v1.3 if you have it. If you bought what is in that EBay listing, it has 1.3 on a sticker, that's what I'm after.

Yep, it's 1.3. I'll pull the PROM and get a .bin for you.

feipoa wrote:

Your story of decreased reliably when overclocking per different day is familiar; you might want to swap out your Am5x86 for another one. It may still be CMOS timing related, although, if your Cyrix 5x86-120 is stable, then it is probably not due to CMOS timings. Usually the Cyrix 5x86-120 is harder to get stable with fast timings than is the Am5x86-160.

The Cyrix 5x86 is rock solid at 120MHz when using FPM with the RAM timings maxed out. If I use EDO RAM, the READ timings need to be dropped to "FAST" instead of "FASTEST" regardless of CPU used (I have three IBM 5x86c) and FSB speed or it will often lock up when ALT-CTL-DEL us pressed--otherwise stable.

It was very strange indeed, the other day I had no problems at all with the exact same Am5x86 at 160MHz but yesterday it could barely even boot into DOS. I'll try my Kingston module and see what happens...

feipoa wrote:

I too had issues with the ALi board and fastest timings. But I really need to retest this with only 256K cache. The more cache you have, the less likely the fastest timings will work.

What speed is your RAM? Try 60 ns.

All sticks of RAM tested were 60ns, using only a single stick.

feipoa wrote:

Is it only the Cyrix 5x86-120 which has issues with EDO and fastest? Does the CYrix 5x86-100 have issues with the fastest EDO setting?

Yes- Cyrix or AMD, 33MHz of 40MHz FSB, I cannot use the fastest read timings when EDO is installed otherwise CTL-ALT-DEL will usually lock up and corrupt the CMOS settings. If I can resist the habit of pressing CTL-ALT-DEL to reboot, it's perfectly fine though... so I left the timings as fast as possible for the benchmarks.

feipoa wrote:

I'm really surprised the Quake and DOOM results were practically identical. An 8 MB/s increase in cache speed should really have showed up in the benchmarks. Something is up. Could you also run cachechk -d -w -t4

This will check the RAM write times as well.

I certainly can. I'll try this and report back.

Thanks!

I recall with the Biostar MB-8433UUD board that with a 40 MHz FSB, I cannot use the fastest cache or RAM timings when EDO RAM is installed. You are learning first hand that EDO and 486's aren't mated well.

I first discovered that EDO RAM in the MB-8433UUD was not stable at 40 MHz when doing 50 GB network backup transfers. It was really irritating because the backup would BSOD half way through. There is something about the EDO which required me to use slower cache timings to achieve stability. Using FPM solved that problem and let me run everything fast.

I have to ask, what does one do with 50 gigs of data on a 486-class machine??

Well, whatever it is, I guarantee it's not going to happen quickly.

It was an HTTP file server (SCSI), among other things. Uptimes were usually around 6 months. No BSODs. It was from a time before the internet exploded with cheap online storage.

What was frusturating is that when I reduced the cache timings from 2-1-1-1 to 2-2-2-2, the system was almost stable, then BSODed around 26 GB transferred. Tried 3-1-1-1, crashed 30+ GB transfered. Tried 3-2-2-2 (the slowest), and it did not crash at all during the transfer. It crashed a few transfers later. At which point, I ripped the EDO out of them and ran FPM without issue. As you've seen, there is no practical advantage to EDO, even if it was stable.

Here's the STB Nitro GX v1.3 video BIOS.

Also, I tested the write speeds with CacheChk.

FPM: 54.5 MB/sec
EDO: 81.7 MB/sec

I too am surprised that with such a difference I don't see any change Doom/Quake/PCPlayer performance.

At first I thought maybe I had a PCI 1/2 divisor enabled with the 40MHz FSB which could be holding back those benchmarks, but I'm pretty sure I don't have it enabled. In the manual I attached back in this post (A Tale of Two 486 Motherboards (ALi vs SIS)) I think it's either JP10 or JP11. I forget which--but one of those makes no difference at all, and the other makes a very small difference--so that leads me to believe that one is the PCI clock divider. I'll have to measure the PCI clock frequency with an oscilloscope to make sure.

Thanks for the BIOS. I've saved it.

It really looks to me like your motherboard is adding/subtracting wait states when different RAM is used. What you might want to do now is run those benchmarks again, but with the RAM/cache CMOS settings set to stable values with the EDO RAM installed. I think, then, you'll only want to use FPM. A good first step for stability is to try installing Windows 98, or NT4 without incident.

Had another look at my ALI board and guess what?
It has socketed cache chips but the chips indeed have "Write back" inscribed over the top which is quite unusual IMO, I've purchased a few 486 motherboards this year and got mixed up with another VLB board that has soldered cache.
Anyway I removed the cache chips and they're unusual, every cache chip I've ever seen the legs look like a dull metal like lead or galvanized steel, these ones are shiny like tin and I suspect they're fake.

The motherboard had the jumpers set to 256K and all these cache chips are the standard size of a tag chip, perhaps they would've worked if I set the jumpers to 128k but I'll never know.
I scavenged 256kb of cache from another dead board and the ALI now works with cache!

This is where it gets a bit unusual but maybe this is the way it is with ALI or 486 motherboards with & without cache.

I get better performance with cache disabled than with cache enabled!
Their seems to be an easy explanation, with cache disabled the memory bandwidth in speedsys is 238.5mb/s, L1 cache 173.98mb/s & memory throughput 62mb/s, with cache enabled the memory bandwidth drops to 119.3mb/s, L1 cache to 160mb/s, L2 cache=54mb/s & memory throughput drops to 34.57mb/s

The performance in dos is close but the cacheless benchmarks win every time.

(Cache enabled/disabled)

Quake - 15.4/15.9
Pcpbench 640x480 - 9.6/10.1
Pcpbench 320x200 - 22.1/22.3
3Dbench 2 - 97.6/101.2
Doom2 timedemo 2 48.6/49.3

The board orginally had a single stick of 32meg fpm ram, on the bootup screen it didn't detect edo ram so I thought I'd try a single stick 16meg stick of edo, I might've gained .1fps in Quake seemed to have made stuff all difference.

FesterBlatz: If you're able to, could you please disable L2(external) cache on your ALI board and then set your dram read and write speeds to "fastest" and re-run the benchmarks?

I wonder if thiers something wrong with my board or a common performance penalty with L2 cache enabled on ALI 1487/89 motherboards, I found that not having the dram read/write speeds not set to fastest really afffects the peformance with cache disabled and even with cache enabled and fastest dram settings the performance drops off.

What kind of cache did you use ? Maybe the latency is too high ?