Try ctcm from heise to test this issue.
https://ftp.heise.de/ct/ctsi/ctcm17a.zip

use graphics card in a 386 sx, dx or any newer:
ctcm /vid
for testing 32 bit transfer speed: ctcm /vid /stosd
for testing 16 bit transfer speed: ctcm /vid /stosw
for testing 8 bit transfer speed: ctcm /vid /stosb
you also may try different video modes since some chips have abysmal performance in some modes

I used this tool to see that some chips are installed in a 32 bit bus but do not use more than 16 bit to the graphics chip, like VLB Cirrus 542x - and to see that my ELSA Winner 1000 EISA really uses 32 bit.

Possibly only tangentially relevant, but some VGA chips only had an 8-bit internal memory bus, AFAICR.

Looks like there is partial support for 16-bit access on WD90C00. I have one of these cards in my 286 and it benched at 3MB/s in mode 13h

noshutdown wrote on 2025-03-07, 03:16:

most isa vga cards are 16bit, which are believed to be faster than the 8bit ones. however, some chips have only 8bit bus so they are faked 16bit cards and performance should be the same as their 8bit variants, and some have been proved by experiments.

Maybe yes, but there's something to consider:
- 16-Bit ISA models may have interleaved 16-Bit wide VGA BIOS, which is faster (unless a copy of VGA BIOS was being copied into shadow memory)
- 16-Bit ISA slot has contacts for zero wait state operation
- 16-Bit ISA cards can run at higher bus speed, no matter whether 8 or 16-Bit transfer

So in theory, an 8086 XT and an 8088 XT with 16-Bit slots (but with the limits of XT architecture; limited IRQs, limited 1MB address range) can be much faster than a normal 8088 PC/XT.
The 8-Bit PC bus is the real bottle neck, even to the 8088/8086 processor.
8-Bit VGA chips might be more capable on other bus systems, such as 16-Bit MCA.

Edit: What also comes to mind, games like Doom used 8-Bit writes.
Not sure how much a 16-Bit chip might have made a difference here. 😕

Edit: What also would been nice: If the VGA BIOSes on these early VGA cards would have made heavy use of 80186/80286 or V20/V30 optimization.
The Windows 3.0 VGA driver did it right, I think, but it probably used 80286 instructions by accident merely.

noshutdown wrote on 2025-03-07, 03:16:

realtek 31030 and 3105 are also highly suspected, simple because their 16bit isa bus higher data pins are often left unconnected. however 3102 and 3103 can not be convinced from circuitry.

There are multiple chipsets for RTG3105 - an ordinary RTG3105/3105e card will have traces that lead to 16-bit ISA pins, but a RTG3105iEh card will never have 16-bit pins connected except for power delivery. RTG3102 is definitely the slowest VGA card in my entire collection, while the RTG3103 from my limited testing has similar speeds to my RTG3105e card.

Another example of a chipset where cards do not have 16-bit pins connected are those that use the HMC HM86314Q chipset.

Here's a picture of a low-cost model of RTG3105IEH.
https://vgamuseum.ru/gpu/realtek/realtek-rtg3105ieh/

It's astonishing that it has almost *everything* inside that chip, including the RAM DAC.
Except for video RAM (512 KB) and a lone crystal.

Such a cheap all-on-one VGA chip must have been a great upgrade to users in third-world countries,
who previously had been used to CGA graphics and 8088 PCs. Like former C64 users here in Germany. ;)

DEAT wrote on 2025-03-07, 14:06:

There are multiple chipsets for RTG3105 - an ordinary RTG3105/3105e card will have traces that lead to 16-bit ISA pins, but a RTG3105iEh card will never have 16-bit pins connected except for power delivery. RTG3102 is definitely the slowest VGA card in my entire collection, while the RTG3103 from my limited testing has similar speeds to my RTG3105e card.

Another example of a chipset where cards do not have 16-bit pins connected are those that use the HMC HM86314Q chipset.

Oh really? I doubt that. I have observed multiple 3105/3105e(not the 3105ie or ieh) cards without fingers for the high 8bit data pins, even those with fingers lack connected traces.
Only one or two cards that i saw show ambiguous traces connected to these high 8bit data pins, but I cannot determine if they lead to the video chip or serve as camouflage (similar to my uncertainty with ET3000 and TVGA8800 cards). At least, these traces do not go straight to the video chip like the ones on 3106 cards do.

i also attempped to find out the difference between 3105 variations, these are all that i found:
3105 and 3105e are regular chips with external ramdac and bios, 3105 cards usually have a 40mhz crystal, while 3105e cards usually have 48mhz.
3105i has ramdac and bios integrated, with a small crystal left, and i can't identify its frequency.
3105ie and 3105ieh both have ramdac, bios and crystal integrated, , and i dunno whats the difference.

darry wrote on 2025-03-07, 12:47:

Possibly only tangentially relevant, but some VGA chips only had an 8-bit internal memory bus, AFAICR.

so far the only vga cards that i know to have 8bit dram width are cirrus 5320 and 5401(avga1 bought from acumos), they are both max 256kb cards and do not support high resolution 256color modes, so 8 bit dram width is actually sufficient for that, and the 5401 is actually quite fast in vga tests.

It has a cool demo application!

https://www.youtube.com/watch?v=fk0EZG1bTg0

Edit: There's also a Windows 3.0 and VESA VBE driver. It seems I already talked about the card years ago.
Re: Cirrus Logic 5320 win3.1

The cirruswd.zip contains a 640x480, 800x600 and 720 by *something* driver.
It has to be installed in system.ini manually, I think.
I wonder what the 720 resolution driver is like. If it is palettized or not etc.

PS: The later Cirrus Logic GD5428 (C-Bus) and GD5430 (PCI) were notable as early GUI graphics accelerators on PC-9821 platform.

noshutdown wrote on 2025-03-10, 01:33:

darry wrote on 2025-03-07, 12:47:

Possibly only tangentially relevant, but some VGA chips only had an 8-bit internal memory bus, AFAICR.

so far the only vga cards that i know to have 8bit dram width are cirrus 5320 and 5401(avga1 bought from acumos), they are both max 256kb cards and do not support high resolution 256color modes, so 8 bit dram width is actually sufficient for that, and the 5401 is actually quite fast in vga tests.

There are others.

http://www.s100computers.com/My%20Syste ... GA9000.pdf

darry wrote on 2025-03-10, 06:02:

There are others.

The attachment Screenshot_20250310-020109-924.png is no longer available

http://www.s100computers.com/My%20Syste ... GA9000.pdf

it can work with 8 bit of ram, but that doesn't mean it has only 8bit of ram width. even the trident8900 can work with 8bit 256KB of ram, but it can support 32bit ram width at max.
ram width depends on amount of ram data pins, not the minimal amount of ram width that can get it running.

noshutdown wrote on 2025-03-10, 09:33:

darry wrote on 2025-03-10, 06:02:

There are others.

The attachment Screenshot_20250310-020109-924.png is no longer available

http://www.s100computers.com/My%20Syste ... GA9000.pdf

it can work with 8 bit of ram, but that doesn't mean it has only 8bit of ram width. even the trident8900 can work with 8bit 256KB of ram, but it can support 32bit ram width at max.
ram width depends on amount of ram data pins, not the minimal amount of ram width that can get it running.

Agreed, but card implementations with minimal configurations likely do exist, possibly even non upgradable ones.

I'm confused with the Paradise VGA Plus 16 (PVGA 1A) is it 16bit or 8bit (faked 16bit)

September 1989. 256KB, 80NS ram, 40Mhz RAMDAC.
In the manual it says
"In addition to a 16bit wide video memory interface the Paradise Plus 16 can present a 16bit wide video BIOS data path, operating twice as fast as an 8bit Vga board".
It also mentions a feature called AUTOSENSE, "a sophisticated hardware and software system that runs tests and automatically enables 16bit BIOS mode". ??

If I force this card with the jumpers to 8bit mode, the bios screen takes about 2 seconds to draw each new page. In 16bit mode the bios screen updates almost instant.
This card also works in 640x400x256 mode in Dos and Windows.

Riikcakirds wrote on 2025-03-12, 20:38:

I'm confused with the Paradise VGA Plus 16 (PVGA 1A) is it 16bit or 8bit (faked 16bit) […]
Show full quote

I'm confused with the Paradise VGA Plus 16 (PVGA 1A) is it 16bit or 8bit (faked 16bit)

September 1989. 256KB, 80NS ram, 40Mhz RAMDAC.
In the manual it says
"In addition to a 16bit wide video memory interface the Paradise Plus 16 can present a 16bit wide video BIOS data path, operating twice as fast as an 8bit Vga board".
It also mentions a feature called AUTOSENSE, "a sophisticated hardware and software system that runs tests and automatically enables 16bit BIOS mode". ??

If I force this card with the jumpers to 8bit mode, the bios screen takes about 2 seconds to draw each new page. In 16bit mode the bios screen updates almost instant.
This card also works in 640x400x256 mode in Dos and Windows.

The PVGA 1A supports 16-bit access to video memory if the memory interface is configured in odd/even mode (text mode, CGA graphics) or in chain-4 mode (standard 256-color graphics mode), but not if the memory interface is configured in native EGA mode (16-color graphics modes). The chip-to-memory interface is fixed at 32 bits, just as it is on the original EGA/VGA hardware. The limited ability to support wide bus transfers in 16-color modes is widespread on second-generation VGA chipsets, and not a unique property of the PVGA chip.

The performance difference you observe in the BIOS screens are likely due to the 8-bit vs. 16-bit BIOS access and possibly your mainboard throwing a lot of wait-states on 8-bit memory access.

Riikcakirds, in very simple words it's like this:

The Paradise cards with an PC bus connector have an single, 8-Bit wide EPROM containing the VGA BIOS in one piece.
The PVGA1A chip is connected in 8-Bit, too.

The Paradise cards with ISA bus connector have two BIOS chips, each being 8-Bit wide.
The VGA BIOS is being split into two, with bits&bytes stored in interleaved form (odd/even).
On an AT, they can be accessed naturally in 16-Bits.
On an PC, the stream of information of the two chips is being combined into an 8-Bit compatible format.
How exactly this works and whether the PVGA1A is involved, I would have to check.

If shadow memory for video BIOS is enabled, a copy of the VGA BIOS is placed in RAM (at the location the ROM had been).
Then, it doesn't matter anymore if VGA BIOS was stored in one or two chips.
The same is true if a "rambios" for PVGA1A was being loaded from the utility disk.
It installs a memory resident copy of a Paradise VGA BIOS in RAM.
It can either be a version stored in the executable or the version onnthe graphics card.

PS: Some later ISA bus VGA cards do have a single 8-Bit EPROM, too, instead of two EPROM chips.
It might be that here, the VGA chip or one of the support chips do provide an interface to 16-Bit bus somehow.

mkarcher wrote on 2025-03-12, 21:10:

The PVGA 1A supports 16-bit access to video memory if the memory interface is configured in odd/even mode (text mode, CGA graphics) or in chain-4 mode (standard 256-color graphics mode), but not if the memory interface is configured in native EGA mode (16-color graphics modes). The chip-to-memory interface is fixed at 32 bits, just as it is on the original EGA/VGA hardware. The limited ability to support wide bus transfers in 16-color modes is widespread on second-generation VGA chipsets, and not a unique property of the PVGA chip.

Picture didn't show in last post, this is the card:

So 16 bit access to video memory in text, CGA and VGA modes.
One thing i don't understand in the datasheet is when it says "the I/O data path is 8 bit". As it also says before this it "decodes all necessary I/O to AT bus in 8 or 16bit data path modes".

mkarcher wrote on 2025-03-12, 21:10:

The performance difference you observe in the BIOS screens are likely due to the 8-bit vs. 16-bit BIOS access and possibly your mainboard throwing a lot of wait-states on 8-bit memory access.

If I shadowed video bios I guess this is void and it will access faster than 16bit bios mode speed, even if set card to 8bit mode but video shadow on.

Riikcakirds wrote on 2025-03-13, 18:41:

One thing i don't understand in the datasheet is when it says "the I/O data path is 8 bit".

ISA cards support two different kinds of access: Memory access and port I/O. The datasheet intends to say that the PVGA1A chip is not able to take 16 bit at once for a port I/O cycle. 16-bit access is only supported for video memory access (in suitable video modes) and BIOS ROM access.

Riikcakirds wrote on 2025-03-13, 18:41:

As it also says before this it "decodes all necessary I/O to AT bus in 8 or 16bit data path modes".

The edition of the data sheet I look at says: "decodes all the necessary memory and I/O addresses to interface with the AT bus (or Microchannel bus), in 8 or 16 bit data path modes." This means that the PVGA1A needs no external chips that find out whether an I/O or memory address is targetting the VGA card, but the PVGA1A has all the address recognition (called "decoding") logic built into that chip. The PVGA1A supports both the classic ISA as well as the "modern" IBM Microchannel protocol. 16-bit support is optional and can be enabled by the BIOS in PVGA register "PR1", othewise 8-bit operation is fully supported, too.