Not terribly hard to pull off - a 430TX board running at 83FSB (with a 10.375Mhz ISA bus, as they're locked at 1/4 of PCI clock) can get around 36.8FPS with any card that hits the maximum speed of the ISA bus. A 25Mhz ISA bus to achieve that seems sub-optimal in comparison.

It's hard to get on a 486 or 386-486 hybrid with only 16-bit ISA slots.

I've had 22.5Mhz on a 386sx/486 upgrade motherboard. Video did work at 25Mhz but either the IDE controller or CF card didn't work.

There's other factors that add to the 35FPS... Faster FSB alone gives a chunk of performance as well as faster cache.

If the game slows down to 10FPS in some areas but speeds up to 30FPS when you face a wall, it's better to work on the low frame rate cause (cache size, speed) instead of more ISA clock. IMO

MikeSG wrote on 2025-03-15, 07:57:

It's hard to get on a 486 or 386-486 hybrid with only 16-bit ISA slots.

I doubt this - a good chipset like UMC will get much better results:
Re: ISA overclock on BX and Via mobos

That YouTube video only shows achieving 21.0FPS at 10Mhz ISA bus with a SiS471 chipset and a AMD 5x86 running at 3x50, which is significantly worse than my 31.2 FPS result with a UMC UM8498F chipset and a AMD 5x86 running at 4x40 FSB at the same ISA bus speed, and that mobo I'm using is a PC Chips M912 v1.7 with fake cache. Not to mention that the UM8498F was the preferred chipset for late-era ISA-only 486 boards, most of those completely lacking L2 cache:
https://theretroweb.com/motherboards/?page=1& … %5D%5Bsign%5D==

While ISA mach32 can hit the top ISA speed with mode 13h, "Mode X" is at about 10% slower than Cirrus Logic GD542x, WD90c3x and Headland HT216 which all hit top speeds, at least with my 430TX and AMD K6-3+ (6*83) setup where I've done a lot more tests on:

(those yellow results for Quake's vid_mode 1 are minor visual glitches with UI elements, since both cards are sensitive to higher ISA bus speeds)

The SiS 486 chipsets have poor ISA performance - I immediately discarded testing with my SiS 496/497 board (Lucky Star LS-486E) as the results with a GD5434 were atrocious:

1SiS496/497
2==========
3AMD DX5-160 (10Mhz ISA)
4doom 17.830
5
6AMD DX5-160 (13.3Mhz ISA)
7doom 20.944

You said the test at 10Mhz ISA was originally done on the 430TX motherboard at 83Mhz FSB....

Maybe you could redo the tests on the UMC UM8498F if you get the same as the 430TX motherboard on a 486 with no cache... that doesn't sound right

He gets: (https://youtu.be/t7Wwx1S-e6E?t=1106)
35.7 @ 3x50Mhz, 25Mhz ISA
30.5 @ 4x40Mhz, 20Mhz ISA
30.4 @ 5x50Mhz, 16.6Mhz ISA
30.4 @ 3x50Mhz, 16.6Mhz ISA

The SIS 496/497 scores are accurate for a 486 at that ISA clock.

At 10Mhz, getting the same score in Doom for a CL-GD5424 and CL-GD5434, means the card isn't utilised as it would be at 16Mhz+ ISA. So those tests can't really tell the cards apart in Dos.

DEAT wrote on 2025-03-15, 22:13:

The SiS 486 chipsets have poor ISA performance - I immediately discarded testing with my SiS 496/497 board (Lucky Star LS-486E) as the results with a GD5434 were atrocious:

Is there a place where the ISA speeds are ranked for different 286/386/486 chipsets?

Many ISA video cards can handle just fine 25MHz and higher frequencies.
Especially true for later model ISA cards based on VLB class video chips.
These chips are designed to operate at 40MHz and higher. Some of them can do 60 and 66 MHz.
So, nothing unusual really.

For example, here is a quick run captured on video showing Cirrus Logic GD-5434 ISA graphics card by Diamond in Shuttle HOT-433 revision 4.
ISA bus at 33.33MHz, FSB at 66.66MHz, CPU at 133MHz (2 x 66.66).
Doom = 41.63 fps.

The achieved framerate can be increased further by optimizing EDO and L2 cache wait states. Didn't push too much in the brief test today. It think ~43 fps is totally achievable.

Some visual artefacts can be observed in the video. They are related to L2 cache, system memory, and video memory.
Didn't spend time to curate them for glitch-free result but posted multiple videos on the subject in another thread that confirm it.

ISA bus running at 40MHz can be demonstrated as well, but that is a lot trickier because of the local storage controllers - very few can handle frequencies above 25MHz. It takes significant effort to get them past ~30MHz.

Several other motherboards can deliver similar results.

Further overclocking the CPU has diminishing results as the slow place is the ISA bus.
For example, running the CPU at 200MHz (3 x 66) and the ISA bus as 22MHz produces around 32 fps in DOOM.
So, a balanced formula is required - fastest possible ISA bus at the expense of other factors, but without completely compromising them.

---

@douglar
I examined that area quite a bit, but don't have the information in one place.
My signature points to a large thread but data is scattered throughout.
Brief summary:
286: ISA bus can run stably at 30MHz, 0 wait states, in 1-2 selected motherboards
386, ISA bus can run stably at 25-27.5MHz, 1 or 2 wait states, in 2-3 selected motherboards
486, plenty of options for 25+ MHz, handful of options for above 30MHz.

In general ISA video cards are not a problem. The limitations are imposed entirely by local storage controllers.

I would like to see the bus signals on an oscillograph.
Because I wonder if they're still clean and stable at a frequency this high.
The VLB bus had issues being overclocked, too, eventually.
The problem with high clock rates is that the mainboard layout and the layout of the expansion cards must do handle the frequency.
Long traces with sharp edges cause reflections or higher inductivity (like a coil would cause).
In worst case, parts who aren’t supposed to oscillate then do oscillate.
Things like unwanted impedance transforming happens through those long traces (they are like inductors/coils).
I think this RF aspect if often being overseen. It's not just the ICs who have a speed limit, the PCB design has an effect, too.

pshipkov wrote on 2025-03-17, 00:21:

@douglar I examined that area quite a bit, but don't have the information in one place. My signature points to a large thread bu […]
Show full quote

@douglar
I examined that area quite a bit, but don't have the information in one place.
My signature points to a large thread but data is scattered throughout.
Brief summary:
286: ISA bus can run stably at 30MHz, 0 wait states, in 1-2 selected motherboards
386, ISA bus can run stably at 25-27.5MHz, 1 or 2 wait states, in 2-3 selected motherboards
486, plenty of options for 25+ MHz, handful of options for above 30MHz.

In general ISA video cards are not a problem. The limitations are imposed entirely by local storage controllers.

That's interesting stuff but I have the feeling that there's going to be some variability from board to board within the same model and from chip to chip within the same product line, like you'd expect for any over clocking effort. Would certainly be interesting to know which ones have a higher ceiling though.

Is there different ISA performance between different chipsets running at the a normal frequency or is ISA so simple that there's not a lot of variation?

Ok, captured another video - this time 200MHz CPU (3 x 66), 33.33MHz ISA, same video card CL GD-5434. For completeness.
Doom reaches 46.5 fps.
Not bad at all.
A bit more display degradation. I am not doing great job at properly showcasing this. May replace the videos later if find the motivation to curate RAM/L2_cache. Anyhow.

@douglar
From practical experience I can attest that same model motherboards exhibit the similar qualities in terms of stability and overclocking.
To your other point - i can provide some compact notes about good/best chipsets for the different hardware classes if there is interest. Not sure if this thread is the right place for it, so hesitant to spam.

@Jo22
Which signals more specifically? On ISA pin 20? It is fine really. No big story there.
I feel that your notes are mostly concerning designers/integrators of such hardware but even mere users can experience it first-hand if on a mission to find working set of level 2 cache chips that work well at high frequencies and tight wait states, or when testing bunch of motherboards based on the same chipset but exhibiting significant differences in performance and scalability on overclocking (with some small exceptions here and there).

---

The source of variability, especially for overclocking, comes from CPU, L2 cache, RAM, local storage controllers, FPU (386 only), keyboard controllers (286, 386), crystal oscillators (single most extreme 386 case), in the listed order.
Stability/overclocking varies greatly between the different motherboard models, even if they are based on the same chipset. This is an indication that circuitry, layout, assembly, integration matter. As you mentioned above.
So that's that.

About the ISA bus - it kind of has a stigma, undeservingly.
Most people conflate what was ISA in IBM 5159 or close cousins (slow, stuck at 4 to 8 mhz) with what was in later chipsets/motherboards.
While actually it is on the same level of implementation quality with everything else that goes into a motherboard from a given time period - as good as AGP, PCI, VLB and so on circuitries. It carries over its inherit design limitations but that is a different story. So, we should not be surprised to see it ticking at 33-40MHz.
The actual limitations in terms of high ISA bus frequency are elsewhere - first - what tools the motherboard offers to get to a fast ISA bus while keeping the rest of the components within their working range, and second - what is inserted into the ISA slots - for example - early Realtek video card won't cut it past 15MHz and so on.

I'd be surprised if most motherboards can't handle high FSB & ISA clocks. The chipset itself though, especially 386... if there's newer made chips available they'll certainly handle higher speeds.

I swapped a 20-25Mhz specified Chips 82C836 for a new one and it did 50Mhz like it was OEM. 25Mhz ISA, but only the VGA card ran. Not IDE.

pshipkov wrote on 2025-03-17, 04:33:

@Jo22
Which signals more specifically? On ISA pin 20? It is fine really. No big story there.
I feel that your notes are mostly concerning designers/integrators of such hardware but even mere users can experience it first-hand if on a mission to find working set of level 2 cache chips that work well at high frequencies and tight wait states, or when testing bunch of motherboards based on the same chipset but exhibiting significant differences in performance and scalability on overclocking (with some small exceptions here and there).

Hi, what I'm thinking of is not so much the digital technology, but the signal quality/stability on the PCB at higher frequencies.
Not sure how to put this into words.

It would be easier if knowledge about building crystal radios receivers or shortwave radios was there.

In short, if you're working with higher frequencies (shorter wave lenghts),
the disctances of the chips and the lenght of the traces should be shorter.

An old 80286 mainboard from 1984 was built with slow frequencies in mind.
There had been longer traces, which were fine up to a certain frequency (let's say 10 MHz).

But if you would run such a large board at 100 MHz,
even if the chips were specced for 100 MHz, then the signal quality would start to become bad/unstable.

Not because of the chips, but because the long wires act as inductors and because they would cause reflections and impedance issues.
The electric contacts of the ISA bus connector would be too big, not fine enough.
Grounding lines would be required betwern each data pin.
The layout of the traces would not up to it handling the high frequencies.

There would be more ground areas needed, more caps to block spikes on the signals.
More layers in the PCB to provide better shielding and better grounding.

In a nutshell, that's what I meant.

The ISA bus or AT bus can technically run as fast as an 80286 can, so 25 MHz is alright.
However, the PCB layout has to take this operation frequency into account.

If the 80286 mainboard is a compact motherboard,
with well designed PCB layout, then it can handle these 25 MHz just fine.

The ISA cards, too, if they were made highly integrated (=few, short traces).

Again, it's hard to put into words.

A bit of radio homebrewing background is needed, I suppose.

In short, as an analogy, a shortwave radio or AM radio can be built on a kitchen table using big parts and long wires.

Doing same with a microwave radio wouldn't work,
because the long wires and big components would start to show weird effects on such high frequencies (small wavelenght).

While, digital technology isn't same as analogue radio technology, it shares some common things.

If you operate a bus at high frequency, you will cause rapid on/off cycles.
You will cause rising and falling edges. And ICs have duty cycles, too.

If you now combine this with long traces and not correctly chosen blocking caps for the opetating frequency,
weird things like oscillations/reflections and harmonics can occur.
Or noise, in short. The signalpath nolonger is clean between the components.

Again, it's hard to put into words.

I also don't say that ISA can't run at higher frequencies per se. It's just the 80286 front side bus, after all.
I merely meant that there's more to consider than just the specs of the chipset, the cache chips, the DRAM and so on.

The physical medium, the circuit board, is also to consider.
The issues with VLB bus at higher speeds can be used for comparison, maybe.

These concepts are well understood at a basic, surface level - the way we casually talk about them here.
It's actually a much more complex topic for folks developing products that are impacted by these factors.
I actually had bunch of signal processing classes in university but never got to practice any of it outside the labs at the time.
That said, I got what you were trying to say in your first post and tried to respond, but I guess my explanation fell short.
Basically, my point is that it's impossible to figure out how significant these factors are just by watching how some old hardware performs.
But after spending a lot of time with extreme overclocking on this kind of hardware, I've started noticing many subtle nuances that hint at the subject here.
I can tell that PCB design isn't as critical as the main components like chipsets, processors, memory, L2 cache and so on - which is basically what I was trying to get across in my first reply.
The clearest example where you really notice this stuff is when you're sorting through L2 cache chips for an overclocked setup and tweaking those tight BIOS timings.
So yeah, we're definitely thinking along the same lines.

Mike, pass a visual reference for your board. I remember you mention it in the SXL2 interposer thread but still don't know which one it is. Or if you posted somewhere in the thread there i either didn't see or forgot about it. I have almost 0 experience with 386SX stuff, other than 2-3 boards like Alaris Cougar and Matra 486SLC2.

pshipkov wrote on 2025-03-18, 06:50:

Mike, pass a visual reference for your board. I remember you mention it in the SXL2 interposer thread but still don't know which one it is. Or if you posted somewhere in the thread there i either didn't see or forgot about it. I have almost 0 experience with 386SX stuff, other than 2-3 boards like Alaris Cougar and Matra 486SLC2.

It's a Seritech 386CS, looks similar to the Biostar MS-1320/25C: https://theretroweb.com/motherboards/s/biosta … mb-1320-25c-b-5

In my experience/theory the length of the trace is not that big of a deal when working with 10-100Mhz. The longer it is, the more it typically has bends (creates reflectance), and runs by GND (creates capacitance/load). Inductance may be there, but normally the board has a guard ring/GND shield around the edges of the board.

In many 386-486 boards the high speed lines run in straight lines, separate to GND by some margin.

So I put forth the argument that the chipset silicon (nm process, internal transistor quality) determines a major part of the FSB speed, and ISA, bus, cache speed.

MikeSG wrote on 2025-03-16, 06:18:

You said the test at 10Mhz ISA was originally done on the 430TX motherboard at 83Mhz FSB....

Maybe you could redo the tests on the UMC UM8498F if you get the same as the 430TX motherboard on a 486 with no cache... that doesn't sound right

At 10Mhz, getting the same score in Doom for a CL-GD5424 and CL-GD5434, means the card isn't utilised as it would be at 16Mhz+ ISA. So those tests can't really tell the cards apart in Dos.

Done, though I grabbed a CL-GD5402 instead of a GD-5424 since I decided a few days ago to follow up on this myself and only checked this thread now. All tests were done with running at 3*50FSB, with two 16MB sticks of 60ns EDO RAM, a Pine PT-606G multi I/O card and a 1GB Transcend CF card which is partitioned for 504MB. I don't have a multi I/O card that works with 25Mhz ISA bus. All Doom tests were done with screenblocks 10 (consistent with both the Youtube video in OP and pshipkov's videos) and shareware DEMO3. The fdoom and fdoom13h tests were done with FastDoom v1.1.1, Quake tests with v1.06:

All of the results marked orange had random glitch pixels all over the screen. The yellow doom+fdoom results of the Headland HT216 on 8.33Mhz ISA were glitches that only happened in the HUD, so it didn't affect gameplay that badly. The yellow wolf_286 results for ATI mach32 VRAM are because of glitchy palette cycles with fade ins and fade outs.

The S3 928 (Number #9 GXE) had tapped out at 12.5Mhz where it was heavily glitching in everything, same with the Tseng ET4000AX at 16.67Mhz. The Oak OTI087X (with 32-bit memory bus width) had uniquely turned into an EGA card at 16.67Mhz. My other S3 928, a Diamond Stealth Pro and my OTI087 refused to POST on this board. I didn't try my other OTI087X cards to see if any of them were stable at 16.67Mhz, nor did I test any of my three WD90C30 cards. The glitchy output for Mode X on the Headland HT216 did not become more intense apart from the 8.33/10Mhz difference, which makes me inclined to believe it's a bad implementation on the chipset rather than a bus speed scaling issue.

I did a double check for both the Trident 8900D and C&T F65545 at 16.67Mhz, and they do have regressions in performance.

douglar wrote on 2025-03-16, 13:34:

Is there a place where the ISA speeds are ranked for different 286/386/486 chipsets?

I've been accumulating a bunch of random benchmarks with what I have sitting around, but I don't have that many motherboards to test from that era - I only have a single 386DX motherboard (which uses a Symphony Haydn chipset), since I don't have a specific need for more than one working 386DX mobo.

That said, I have a lot more dead/dysfunctional 286 mobos than I have working ones.

EDIT: I do also have a hybrid 386/486 OPTi495SLC mobo, but the 386 socket doesn't work (it was unpopulated and I've been meaning to recheck my solder job, but it's very low priority) and that chipset has horrible uncached memory performance compared to the UM8498F. I'll upload the screen captures of SpeedSys from my OSSC when I'm back at my home PC.

Bizzarely fast RAM speed on that UMC board. Looks like it was tuned to fully utilise EDO DRAM.

And you're right about the CL-GD5402/24/29 being similar to the CL-GD5434.

A lot of 486 boards are going to have slow memory speed.

I work with mostly 386SX & DX so memory speed is always on the floor. More Mhz does make it faster, but low FPS scenes are always low w/o cache.

DEATs numbers confirm once more that ISA bus is the main perf limit. Thanks for sharing. Good stuff.
If i understood correctly, you are testing with Am5x86 at 150MHz (3x50), correct?
If so, then your results are significantly better than where the youtuber was in his video.
In fact, these are impressive numbers for 16MHz ISA bus.

And don't worry about not having more 386 motherboards.
You got the best stuff already (Symphony Haydn).

@MikeSG
This is actually supposed to be a slow motherboard, but you are squeezing a lot of mileage out of it. That's awesome.