That's Cirrus 5428, right?
I would try to find some software that tells where the linear frame buffer is located - in ISA mode it's below 16 MB, in local bus mode a higher address.
Windows 9x resource manager should tell that, and I think UniVBE/SDD...

See eg. this - LFB at 3072 MB, definitely not ISA.

It's VLB but also do the test above for confirmation

Try to use ASTRA for DOS:

ASTRA - Advanced Sysinfo Tool for DOS

fabiensanglard wrote on 2025-01-06, 23:24:

Speedsy shows a video bandwidth of 8MiB/s which sound inline with ISA 16-bit.

Even theoretical throughput of ISA at 8MHz clock is around 5.3MiB/s, in practice there's almost no way to achieve such throughput.
So unless the ISA bus is overclocked - the 8MiB/s should indicate VLB.

Thank you for all your help, what a gem this community is!

VBETEST.EXE

I was able to find the 1998 version of VBETEST.EXE but not the 2002 version from the screenshot. Nothing was displayed for linear frame buffer. I don't know if this is because my card does not have one or because the 1998 version was too old.

ASTRA

Great tool, it was able to tell me the bus is VLB. I have no idea how it generates this information.

It's VLB but also do the test above for confirmation

You were right @keropi. How did you know? Are 2168 notorious for all being VLB?

Even theoretical throughput of ISA at 8MHz clock is around 5.3MiB/s, in practice there's almost no way to achieve such throughput.
So unless the ISA bus is overclocked - the 8MiB/s should indicate VLB.

Thank you for the correction @vstrakh. My mistake was to start from VLB theoretical max bandwidth (4*33000000/1024= 128,906KB/s) and conclude that 8,365 KB/s was way to low compared to that. I could have figured out from ISA theoretical max bandwidth (8,330 KB/s) which is below my figure (and also way too close anyway, a bus practical bandwidth is always much lower than the theoretical max).

You were right @keropi. How did you know? Are 2168 notorious for all being VLB?

If you pop the case and find CL-GD5424 or CL-GD5425 on the chip, it's VLB. Right up to the 5429. CL-GD5422 is ISA. 5430 was adapted for PCI.

I also found a summary table that corroborate what was said on this thread about the 542, 5421, 5422, 5424, 5425, 5426, 5428, and 5429.

At least up to the CL5428 (and I suspect the 5429, too) is *also* possible to be connected just via ISA alone. I've got one:

That said, I'm in agreement that 8MB/Sec seems to be too high for VGA memory bandwidth on the ISA bus.

megatron-uk wrote on 2025-01-08, 10:28:

At least up to the CL5428 (and I suspect the 5429, too) is *also* possible to be connected just via ISA alone. I've got one:

Yes, 5429 also supports ISA, and even 5434:
Newly made ISA CL-GD5434

fabiensanglard wrote on 2025-01-08, 05:21:

I was able to find the 1998 version of VBETEST.EXE but not the 2002 version from the screenshot. Nothing was displayed for linear frame buffer. I don't know if this is because my card does not have one or because the 1998 version was too old.

That's because you're using the VBE 1.2 implementation from ROM.
For LFB support, you need at least VBE 2.0 - provided by certain versions of UNIVBE.EXE.

Grzyb wrote on 2025-01-08, 13:10:

megatron-uk wrote on 2025-01-08, 10:28:

At least up to the CL5428 (and I suspect the 5429, too) is *also* possible to be connected just via ISA alone. I've got one:

Yes, 5429 also supports ISA, and even 5434:
Newly made ISA CL-GD5434

You could kill the speed of the 5424 by adapting it for ISA, such as what they did with that BOCA card. Pretty sure Cirrus Logic didn't intend for the 5424/5428 to be used that way, but that didn't stop companies from trying to rebuild ISA cards with newer components. BOCA used that PCB for any version of the 542x that they could find and a lot of them ended up in the EU.

Dmetsys wrote on 2025-01-08, 14:28:

You could kill the speed of the 5424 by adapting it for ISA, such as what they did with that BOCA card. Pretty sure Cirrus Logic didn't intend for the 5424/5428 to be used that way, but that didn't stop companies from trying to rebuild ISA cards with newer components. BOCA used that PCB for any version of the 542x that they could find and a lot of them ended up in the EU.

Speed is not all - additional video modes are also important.

Also...
Yes, in DOS, 5420/5422 already hit the ISA bus limit - later chips can't be any faster.
But in Windows, the acceleration features introduced in 5426 - and ehnanced in later chips - can make it faster even on ISA.

Dmetsys wrote on 2025-01-08, 14:28:

Grzyb wrote on 2025-01-08, 13:10:

megatron-uk wrote on 2025-01-08, 10:28:

At least up to the CL5428 (and I suspect the 5429, too) is *also* possible to be connected just via ISA alone. I've got one:

Yes, 5429 also supports ISA, and even 5434:
Newly made ISA CL-GD5434

You could kill the speed of the 5424 by adapting it for ISA, such as what they did with that BOCA card. Pretty sure Cirrus Logic didn't intend for the 5424/5428 to be used that way, but that didn't stop companies from trying to rebuild ISA cards with newer components. BOCA used that PCB for any version of the 542x that they could find and a lot of them ended up in the EU.

You could look at it that way; a reasonably fast VLB part crippled by the ISA bus... or you could look at it the other way; as a more modern, integrated single-chip solution, offering better performance than most of the established ISA chips.

Depends which side of the fence you are on.

When I bought one of these cards new in the early 1990's (93 or 94 I'd guess), as an upgrade for a 512kb Oak card on a struggling 386 SX 40, it was definitely the latter.

The ET4000 was considerably faster than the CL5424 on the ISA BUS. Again, BOCA and other one-off card vendors definitely gimped the 5424. Even the 5422 performed faster in comparison. That being said, anything is better than an OAK card.

Dmetsys wrote on 2025-01-08, 17:42:

The ET4000 was considerably faster than the CL5424 on the ISA BUS. Again, BOCA and other one-off card vendors definitely gimped the 5424. Even the 5422 performed faster in comparison.

I don't know about Boca specifically, but it's clear that properly designed ISA cards, from 5402 (AVGA2) up to 5428, all score the same:
http://vgamuseum.info/images/vlask/bench/quake320.png

That being said, anything is better than an OAK card.

Well, Realtek seems even worse...

Kind of a moot argument to make when you're not using ISA graphics cards with Pentium based systems, unless you have some specific purpose for doing so.

fabiensanglard wrote on 2025-01-08, 05:21:

Thank you for the correction @vstrakh. My mistake was to start from VLB theoretical max bandwidth (4*33000000/1024= 128,906KB/s)quote]

That's not how you calculate the maximum bandwidth on VL. A VL cycle takes at least 2 clocks, so we get down to 66MB/s maximum rate at 33MHz. Furthermore, the CL-GD542x series of chips can not transfer more than 16 bits at once, so we get down to a theoretical maximum rate of 33MB/s. While the VL bus specification might allow for burst writes, 486 processors only do so (if at all) for copying back data from L1 cache if they operate in L1WB mode. As the video memory is uncached, you won't get burst writes at all (and thus I also don't think any VL video chip ever implemented accepting burst writes).

Bus transfer rate is not the only limit, the CL-GD542x series does not implement 0WS cycles on VLB (it's more like 2WS minimum for FSB33, the data sheet specifies a minimum amount of waitstates so the VL bus can't get faster than the video memory interface can take). At 2WS, data rate goes down by a factor of 2 again, which is 16MB/s theroetical maximum that can be accepted by a Cirrus Logic chip on the local bus, if it does not add further waitstates due to the write buffer in the card being full. 8MB/s is actually quite typical for a 5426/5428 on the local bus.

Dmetsys wrote on 2025-01-08, 17:42:

The ET4000 was considerably faster than the CL5424 on the ISA BUS. Again, BOCA and other one-off card vendors definitely gimped the 5424. Even the 5422 performed faster in comparison.

False, even more so when it comes to Windows 3.1 benchmarks. The ET4000 drivers are crap at anything that isn't blitter operations.

Dmetsys wrote on 2025-01-08, 17:42:

That being said, anything is better than an OAK card.

This constant parroting of the same generalised claim I see everywhere without any evidence is remarkable. I don't have access to my raw benchmark data atm, but Doom benchmarks for the OTI067 and OTI077 are within 91% of the GD5434 on a 486SLC-33, the OTI087 and OTI087X are within 97% of the results - that is clearly better than Trident 9000 or S3 911/924 cards which hover closer to 80%. Where the OTI067/077 really suffers is with 800x600x256 and higher on Windows 3.1 - this is especially annoying with an OTI077 that has 256K x4 DIP-20 chips instead of those bizarre 512K x4 SOJ-20 chips.

The OTI087/OTI087X are once again better than the ET4000 in Windows 3.1 by a not-insigifcant margin, but has a weird performance dip in Windows 3.1 at 800x600x256 which then mostly bounces back at 1024x768x256. All OTI087X ISA cards I have are gimped in design - they either have a 16-bit memory bus with 512K x4 SOJ-20 chips which really hurts 800x600x256/1024x768x256 performance, or a bizarre 32-bit memory bus with six 256K x4 DIP-20 chips, two 256K x3 DIP-20 chips and two 256K x1 SOJ-18 chips. Compared to my OTI087 card which has a very clean design, there also appears to be a jumper (and a couple of resistors) missing for enabling 0WS on the OTI087X cards.

I don't disagree when it comes to the OTI037C, though.

Grzyb wrote on 2025-01-08, 18:38:

I don't know about Boca specifically, but it's clear that properly designed ISA cards, from 5402 (AVGA2) up to 5428, all score the same:
http://vgamuseum.info/images/vlask/bench/quake320.png

The funny thing is that last weekend I discovered that VIA boards from Socket 7-onwards have gimped ISA performance compared to other chipsets - the 430TX (especially when running at 83FSB and a K6-3+ at 500Mhz) is overall the best by far, with the 430VX not trailing too far behind (which has better unchained VGA/Mode X scores and has the option to run the ISA bus at PCICLK/3, but it's worse in everything else) and the 440BX at a relatively close third place. The Aladdin V chipset is not too far behind the 440BX, but my GA-5AX rev 4.1 board runs the PCI clock at 33Mhz with 83FSB, and there's no way I can get 120FSB with that board.

Dmetsys wrote on 2025-01-08, 20:19:

Kind of a moot argument to make when you're not using ISA graphics cards with Pentium based systems, unless you have some specific purpose for doing so.

https://theretroweb.com/motherboards/s/ecs-p5st-br
https://theretroweb.com/motherboards/s/apc-hunter-industrial

I have the first board coming my way, should be arriving in a couple of weeks from now.