feipoa wrote on 2021-10-31, 04:55:

1) When using UUD w/1024K at 160 MHz, did the presence of EDO memory improve the benchmark results at all? I see your Speedsys memory results are in the 47's, while normally this is in the 45's for FPMs. Did you have to fiddle around with different EDO modules before you found ones that worked?

This UMC chipset ticks better with EDO RAM. So yes.

feipoa wrote on 2021-10-31, 04:55:

2) I have always had difficulty using EDO on the UUD, albiet they were 60 ns modules. Were 50 ns modules necessary to run at 160 MHz w/2-1-1-1 and 0ws/0ws?.

50ns makes a difference only when base frequency is 60 or 66 MHz.
60ns is perfectly fine with 33, 40, 50 MHz.
Also, good 60ns modules can handle 60MHz just fine. Not sure about 66MHz.

I am surprised to hear you had trouble with EDO modules.
So far this motherboard is just solid with any RAM i threw at it.
4x40 with 2-1-1-1 is 0/0ws is achievable. Takes some combing through L2 cache chips. RAM is never a factor.
So far i managed to do it with both 256K and 1024K L2 cache.

feipoa wrote on 2021-10-31, 04:55:

3) Did you try UUD w/512K double-banked at 3x60 MHz with both EDO 50ns and FPM 50ns?

Yes. Works fine with 3-2-2-2 timings and 1/0ws.

feipoa wrote on 2021-10-31, 04:55:

4) Have you always been using EDO memory while testing the UUD? If so, perhaps this may explain the problematic selection of magic SRAM modules.

At 40/50MHz base frequency i insert whatever RAM module is on top of the stack.
Didn't notice any dependency between SRAM modules stability and type/size of RAM.

---
--- 4x50MHz with 1024Kb L2 cache
---

Some quite interesting results below.

First of all - it was easy to make it happen since the L2 cache chips were already curated.
2-1-1-1 with 1/0 is possible. All other timing settings on MAX.
Tested with 8Mb/60ns and 64Mb/50ns EDO RAM.
8Mb shown in the screenshots and videos below because it offers some time savings between reboots and Quake 1 starts right away = quicker iterations.
64Mb works equally well. At 50MHz base freq RAM is not a bottleneck.

So here is what SpeedSys says.
Left side - L2 cache disabled.
Right side - enabled.
L2 cache performance seems to be close to raw memory access (right graph). Difference seems to be small enough not to show on the limited vertical resolution.
But more importantly - enabling the 1024Kb L2 cache lowers the overall memory metrics.
Resembles the weird case with PC-Chips 918i.

Benchmarks confirm it too. Captured few on video.
Now, the same tests with 256Kb L2 cache and everything else the same - video.
Something is really off.

Few numbers to make it clear (other tests reflect the same ratio):

1                no L2 cache   256Kb   1024Kb
2PCPBench (fps)      27.1      29.2     25.3
3Quake 1 (fps)       20        20.8     19.1

Something is really off with this UUD/UMC/1024Kb business at 4x50MHz.
Somehow 256Kb L2 cache is faster than 1024Kb in this configuration and all other parameters the same.
Not sure what to make out of this.

Those results at 50 MHz and 2-1-1-1 are indeed curious, that is, 4x50 MHz and 1024K results in slower performance than even no external cache. Was DRAM at 0ws/0ws? or 1ws/0ws?

Perhaps every 10th clock cycle is not being read? What happens when you increase the wait state to 3-2-2-2? What happens when you lower the SRAM to 512K? What happens when you run it at 50x3?

It was not just my tests which noticed improved stability with FPM over EDO on this board. I am baffled by your outcome. Not too long ago, I pulled my magic 64 MB FPM module out of my UUD (IBM 5x86c-133/2x) and inserted an EDO stick. It simply could not handle my settings of 3-2-2-2 1ws/0ws. Perhaps I need a 50 ns EDO stick? I think all my sticks are 60 ns. The only 50 ns EDO sticks I have are 168-pin registered ECC DIMMs. My magic FPM stick is also 60 ns. I tryed non-magic FPM 64 MB 60ns stick, but it also could not handle my timings.

pshipkov wrote on 2021-10-31, 07:30:

Something is really off with this UUD/UMC/1024Kb business at 4x50MHz.
Somehow 256Kb L2 cache is faster than 1024Kb in this configuration and all other parameters the same.
Not sure what to make out of this.

You should have two plateaus, one between 0K and 16K and another between 16K and 1024K.
Your graph is like every cache access is a miss--that's why it does nothing but lower the memory throughput.
Since you had to also modify the tag ram when upgrading to 1024K, is it possible it isn't wired right?
Do you get any different results if you lower the cache to 3-2-2-2?

DRAM was at 1/0. otherwise POST does not complete.
If i increase the wait states to 3-2-2-2 - same overall outcome - just further worsened performance.
Didn't run 3x50. Can give it a spin later.

I accumulated bunch of 50/60 ns EDO modules over time. They are not all equal. That's for sure. Looks like your EDO module is lacking.
As i mentioned before - Bio UUD is just so good to me with RAM.

The 1024K setup works well - that's for sure. If you check the previous post for this motherboard where i provide screnshots of the graphs for 4x40 and 3x60 - they all look normal.
Just at 4x50 something is off and that is a motherboard issue.
Yes, it may be cashe misses since the graph is flat after 16kb.

So.
Same with 3x50 or 4x50.
Regardless if 3222 or 2111.
Regardless if 512 or 1024.
With 256k - all good.

Gigabyte GA486AM/S revision 2.20, 2.21 and 2.23 (Feipoa's board).
This was the last 486 motherboard on the short "to examine" list.

Reputable brand. Great looking assembly. UMC chipset with 1024Kb L2 cache.
Acquired it from a member here. There were some issues with it, but managed to resolve them.

And the 2.21 version. Looks identical to 2.20.

All examined motherboards based on this chipset had trouble with bigger L2 cache, especially past the 40MHz base frequency.
So, the question was - can this guy show how it is done ?
Answer is - simply put - no.

The board has issues with different Dallas RTC models. Hint from Feipoa unblocked me. Looks like 12C887+ is preferred.
Jumper settings are confusing.
There are 2 BIOS revisions from year 1995 - A and B. No apparent difference in functionality between them.
On-board IDE controller is early revision and goes up to PIO mode 3 only.
Both L1 and L2 caches in WB mode.
No EDO RAM support.
Had trouble getting bigger volume memory modules to work. Running with 2x16Mb.
Looks like the BIOS does not have proper support for Am5x86 class CPUs. Had to use the simple interposer from this and this posts to enforce L1 WB and 4x multiplier.

Tested with Matrox Millennium PCI and 32Mb 60ns FPM RAM.
Hit an issue with this card and the motherboard. Frequently enough the system will hang upon restart with BIOS beeping "graphics card not recognized". MGA cards are solid and highly compatible. First time encountering such a problem. It is the board, not the card.

--- 160MHz (4x40)

All BIOS settings on max. System is fully stable.
IDE driver is unstable - hangs.

Here is what SpeedSys says:

benchmark results

--- 180MHz (3x60)

BIOS settings on max, except:
DRAM READ WAIT STATE = 1 (best is 0)
CACHE WAIT STATES = 3-2-2-2 (best is 2-1-1-1)
CPU Clock / PCI clock = 1 : 1/2 (best is 1:1), otherwise POST does not complete.
System is stable in DOS but not in Windows. Tried hard to improve on that - without success.

SpeedSys:

Few numbers for reference:
Wolf3D: 127.7 fps (slower than 4x40)
PC Player: 30.1 fps / 9 fps
Doom: 61.7 fps (slower than 4x40)
Quake 1: 19.8 fps
3D Studio R3: 104 seconds (lower is better)

Not great.

--- 200MHz (3x66)

Cannot boot into DOS no matter what. Tried without L2 cache, most conservative BIOS settings, trusted RAM and CPUs.

--- 200MHz (4x50)

No lights.

--- IBM 5x86c at 2x60MHz (going by Feipoa notes)

DOS seems to be ok.
Windows is not happening.

---

Not much more to add for a conclusion.
With some effort this motherboard can work ok at 4x40MHz including, but that's it.

Next up - 486 summary.

Are your results with BIOS rev. A or B? I find it curious that your L2 results at 160 Mhz are 63.85 MB/s, while mine were 59.49 MB/s. Perhaps a difference between L2: WB/WT. I'm guessing I was in WT mode and you in WB.

Note that the 3.3 V CPU core setting on my board measures 3.42 V. I suspect a similar case for your board.

Didn't notice perf difference between the two bios versions.
Don't have good explanation about the speedsys delta.
EDIT: Removed part of the message to minimize confusion around cache modes.

Yes manual says 3.3v but it is 3.45 really.

If you weren't able to run L1 in WB mode, then why are our Quake 1 scores almost the same, 18.1 fps vs. 18.0 fps?

EDIT: Provided the wrong SpeedSys screenshot showing L2 cache in WB mode.

I pulled my board out of storage to compare it with the image you shared, but unfortunately, your image resolution is too low. Could you share a higher resolution image?

For an Am5x86-160 in WB mode and from what I can deduce from the poor resolution, it looks to me like your JP20 is shorted. It should be left open. I also see some differences around JP12, which should be 1-2 shorted and JP11 should be 1-2 shorted. Pin 1 is towards the RTC, not towards the PCB edge.

By chance does your COM1 work?

My board revision is 2.21.

I cannot remember if the jumpers In the picture match what I tested with at the end.
Click on the image for high resolution. EDIT: tweaked the image urls.
At any rate let me know what you have there.
Board is in a antistatic bag but still on the desk. Will cross check tmrw.

Com ports work.

Thanks for ammending the URL link to include the high-res image. The high-res image was previously not showing.

In addition to the 3 jumper settings I noted previously, you should have JP13 open - yours is shorted. You need to short JP14, yours is open.

Not sure if this matters, but I have JP9 shorted to follow Intel settings, while yours is open.

Hard to tell from your photo for JP5, but it looks like 3-4 short and 5-6 short, which is what I have.

I am pretty sure if you match my jumper settings, you'll have L1 WB working. Even after setting up my jumpers originally, I still missed one and L1 WB wasn't working correctly. The one which got me was JP20, but on the silkscreen, part of the '0' in '20' is missing and it looked liked JP22.

Picture does not reflect the final configuration (snapped it at some point during the examination of the board).
You can see that even the voltage jumpers are in default position for 3.3V (3.45).
But the rest is very close to the final configuration.

To clarify:
The important jumper is JP12. It is set to 1-2 for L1 cache in WB mode. Picture shows it in 2-3.
J22 closed for 4V to CPU.
The rest is as shown in the picture, including JP13 and JP20 which are still in closed position.

JP20 state can be overridden by BIOS - L2 cache in WT mode for better performance.
Manual is unclear to me about JP13. It states - closed for "all CPU types" and open for "80486DX4". Left it closed.
The state of these two jumpers does not make any observable difference. Checked again to make sure i am not missing something here. No difference.

If your jumpers match mine and you still don't have working L1 cache, perhaps the board or BIOS has some defect. I ran into this issue once on a PC Chips board and the issue was that the jumper header had a cold solder joint.

Either of you guys have a Rendition Verite based GPU? Curious if one of these bios revisions on the GA486 supports bus mastering? I can use TweakBios to enable it but still feel like there must be a way to mod the bios to allow it permanently, or maybe either the A or B bios version does? Gotta look and see which one I’m using.

Side note, using the Biostar 8433 bios also solves the issue but then you can run into some IRQ issues with other devices.

Chadti99 wrote on 2021-11-07, 23:35:

Either of you guys have a Rendition Verite based GPU? Curious if one of these bios revisions on the GA486 supports bus mastering? I can use TweakBios to enable it but still feel like there must be a way to mod the bios to allow it permanently, or maybe either the A or B bios version does? Gotta look and see which one I’m using.

Side note, using the Biostar 8433 bios also solves the issue but then you can run into some IRQ issues with other devices.

I don't have a Rendition Verite based card. Anything special about it - like some unique and interesting games, or the frame rate is incredible on a 486? I haven't looked into this card, ever. There are DOS utilities to set configuration registers, but it still needs to be loaded at start-up. MODBIN can also set various chipset registers, but you need to look into what is what.

Is there a GLQuake version for the verite, and if so, what kind of frame rates do you get when compared to a Voodoo2?

As for the 8433 and IRC conflicts, you need to play around with the BIOS page that adjusts available PCI IRQ's. There is some wizardry to it that doesn't always make sense, but with enough combinations and trial/error, you can often get it working suitably.

Thanks Feipoa, the Verite’s claim to fame was the first gpu accelerated version of Quake, known as vQuake. You can see 25-33% improved frame rates over software quake between a 5x86160 or POD100 at the same resolution. So with my POD100, on the GA486 board, I go from 29FPS to 39fps, with 16-bit textures and a few other nice rendering features. You won’t see the same performance a Voodoo can muster but you’ll get an arguably more accurate experience mainly due to the lighting being truer to the original software only renderer, and it runs in dos. The gpu is really better suited for a fast Pentium machine but I like one in a socket 3 machine because it still helps and I can run it alongside a Voodoo if I want.

What is a CMOS tap? Is this the bios config screen that allows you to adjust PCI IRQ assignments?

feipoa wrote on 2021-11-07, 22:27:

If your jumpers match mine and you still don't have working L1 cache, perhaps the board or BIOS has some defect. I ran into this issue once on a PC Chips board and the issue was that the jumper header had a cold solder joint.

Ugh, apparently i posted the wrong SpeedSys screenshot.
Took several for the different jumper/bios configurations because of of the unclear purpose of some of the jumpers. Updated the screenshot.
Then replied to your initial question about the different memory timings while outside doing something and not giving it much thought. Apologies.
To confirm - L1 cache is in WB, L2 cache is in WT.

---

@chadti99
I also don't have Rendition cards, also never tried any of them.
"you’ll get an arguably more accurate experience mainly due to the lighting being truer to the original software only renderer"
This sounds interesting. How noticeable is this during normal game session ? Or do you know of A-over-B screenshot references out there ?

Btw, since you have this board there - how is it working for you ?
What numbers you get with Q1, Doom, PCPBench, etc ?

Btw, i tried briefly the UUD bios on AMS - it seemed to work in DOS at least.
Feipoa tried it before me and noticed some issues with Windows (if i remember correct), so didn't drill further.

Chadti99: I've updated my reply to correct the typo. Play with the PCI IRQ assignments in the BIOS. Limiting PCI assignments will also adjust ISA assignments indirectly. What kind of framerate do you get in VQuake at 320x200, sound off, w/Am5x86-160?

I don't see any Rendition Verite's on eBay, so I don't think I'll be playing with this anytime soon. From my understanding, 3dfx offered the best performance on a 486.

pshipkov:It is ok, I make lots of mistakes in typing, normally because kids and wife are yelling at me. I'm sure you have other mistakes in this thread that haven't been caught.

I think I only tried the UUD BIOS on AMS with 60/66 MHz FSB and couldn't get Windows running. I don't think I bothered with 160 MHz in Windows because the AMS BIOS was already working. I did test if the UUD BIOS would allow for the use of 64 MB modules, but it did not - still reads only 16 MB per stick.