Much as we suspect that IBM VGA is the slowest DOS VGA, because everyone wanted to be able to say they were a bit better at least, the slowest Windows accelerated card has to be the OG https://en.wikipedia.org/wiki/IBM_8514 and there's some other compatibles listed in article, anything that claimed 8514 compatibility had the basic acceleration features.

BitWrangler wrote on 2024-12-18, 15:25:

Much as we suspect that IBM VGA is the slowest DOS VGA, because everyone wanted to be able to say they were a bit better at least, the slowest Windows accelerated card has to be the OG https://en.wikipedia.org/wiki/IBM_8514 and there's some other compatibles listed in article, anything that claimed 8514 compatibility had the basic acceleration features.

What about Targa ? That might have been slower than the 8514a

https://www.computer.org/publications/tech-ne … uevisions-targa

noshutdown wrote on 2024-12-18, 12:44:

is it safe to say that the slowest class with acceleration are cirrus5426, oak087 and s3 911?

The OTI087, yes - definitely not the S3 911. No comment on the GD5426 since I don't own one, but I'm going to say no to that based on the GD5429.

Partial accelerators like the Cirrus Logic GD5402/20/22/24, Oak OTI087/087X and Headland HT216/F are all contenders in that they focus on one or two specific things. WD90C31, Weitek W5286 (at least with W5186 drivers) and Avance Logic ALG2101 also don't accelerate everything, but do a better job than those. I have no idea where the Weitek W5186 lands in all of this since I only know of one single card that was released with that chip on ISA bus, the VidTech WinMax. That said, given my research into ISA video cards with Win 3.1 all I'm more certain of is that "framebuffer" and "acceleration" are fairly loose terms - by all accounts people on here would refer to the Headland HT209/D and WD90c30 as framebuffer cards but they certainly do a better job than ET4000, Trident 8900D, OTI077, Realtek RTG3106, ATI VGA Wonder and Ahead V5000B.

Although I dismissed it earlier due to how useless the Graphics WINMARK summary score is with its unusually heavy weighting towards font rendering (and also the tests it focuses on, apart from having separate line drawing scores - memory-to-screen blitter operations are irrelevant for games, screen-to-screen is far more important especially since Program Manager itself uses those) and very valid concerns back in 1993 about benchmark cheating, I do believe that Winbench 3.11 is the best tool for detailed performance information on ISA cards as the full benchmark provides a lot more detail than anything else that I know of.

BitWrangler wrote on 2024-12-18, 15:25:

Much as we suspect that IBM VGA is the slowest DOS VGA, because everyone wanted to be able to say they were a bit better at least, the slowest Windows accelerated card has to be the OG https://en.wikipedia.org/wiki/IBM_8514 and there's some other compatibles listed in article, anything that claimed 8514 compatibility had the basic acceleration features.

i really should have added the constraints: isa vga card, but isa is already mentioned in the topic, and the 8514 is not an isa card.

noshutdown wrote on 2024-12-19, 00:55:

BitWrangler wrote on 2024-12-18, 15:25:

Much as we suspect that IBM VGA is the slowest DOS VGA, because everyone wanted to be able to say they were a bit better at least, the slowest Windows accelerated card has to be the OG https://en.wikipedia.org/wiki/IBM_8514 and there's some other compatibles listed in article, anything that claimed 8514 compatibility had the basic acceleration features.

i really should have added the constraints: isa vga card, but isa is already mentioned in the topic, and the 8514 is not an isa card.

There were 8514a compatible ISA cards.

• Matrox MG-108
• Paradise Plus-A
• Synergy Advanced Technology VGA 8514A COMBO
• ATI Mach 8

And then there were cards based on the lesser known TI 34010 standard:

• Desktop Computing AGA 1024
• NEC Multisync Graphics Engine
• Renaissance Rendition II

This article reviews some of them:

https://books.google.com/books?id=KjwEAAAAMBAJ&pg=PA51

The ATi mach32 can also use the Windows 3.1 generic 8514/a drivers as well (and is also backwards compatible with mach8 drivers) - loses about a third to two thirds of relative blitter performance compared to using mach8 drivers depending on the specific operation.

DEAT wrote on 2024-12-18, 21:53:

The OTI087, yes - definitely not the S3 911. No comment on the GD5426 since I don't own one, but I'm going to say no to that based on the GD5429.
Partial accelerators like the Cirrus Logic GD5402/20/22/24, Oak OTI087/087X and Headland HT216/F are all contenders in that they focus on one or two specific things. WD90C31, Weitek W5286 (at least with W5186 drivers) and Avance Logic ALG2101 also don't accelerate everything, but do a better job than those. I have no idea where the Weitek W5186 lands in all of this since I only know of one single card that was released with that chip on ISA bus, the VidTech WinMax. That said, given my research into ISA video cards with Win 3.1 all I'm more certain of is that "framebuffer" and "acceleration" are fairly loose terms - by all accounts people on here would refer to the Headland HT209/D and WD90c30 as framebuffer cards but they certainly do a better job than ET4000, Trident 8900D, OTI077, Realtek RTG3106, ATI VGA Wonder and Ahead V5000B.
Although I dismissed it earlier due to how useless the Graphics WINMARK summary score is with its unusually heavy weighting towards font rendering (and also the tests it focuses on, apart from having separate line drawing scores - memory-to-screen blitter operations are irrelevant for games, screen-to-screen is far more important especially since Program Manager itself uses those) and very valid concerns back in 1993 about benchmark cheating, I do believe that Winbench 3.11 is the best tool for detailed performance information on ISA cards as the full benchmark provides a lot more detail than anything else that I know of.

ah i thought that bitblt was the core feature of acceleration, and anything with that can be called an accelerator, even if incomplete.
5426 is the first bitblt card from cirrus, 5428 and 5429 are further improved model. 5429 may be a decent average isa accelerator but some blamed that it doesn't have true 32bit bus and thus the vesa version didn't perform very well.
some blamed the s3 924 not performing well for having only 16bit memory width. i am not sure about that but can't find any datasheets, and i thought that 911 and 924 are mostly the same.

From looking through a lot datasheets, the most common stand-out feature to mention is memory-mapped IO, then a command/BitBLT FIFO buffer.

I generally summarise:

Basic accelerators only feature BitBLT & a hardware cursor.
The next best have memory-mapped IO.
The next best have a command/BitBLT buffer, with VRAM being better at higher res/colours. Better drivers.
The next best are 64-bit.

A lot of manuals, especially for the WD90c33, list every minor improvement; buffering, use of off-screen memory, but they're all bested by 64-bit BitBLT'ing.

But what would be a ~2x gain for 64-bit BitBLT'ing is only a 10-20% gain. So the ISA clock becomes a limiting factor.

Those 16-bit memory-bandwidth cards might be the top cards for 286/386sx's. Many drivers for the faster cards only support 32-bit CPUs.

MikeSG wrote on 2024-12-19, 05:48:

From looking through a lot datasheets, the most common stand-out feature to mention is memory-mapped IO, then a command/BitBLT F […]
Show full quote

From looking through a lot datasheets, the most common stand-out feature to mention is memory-mapped IO, then a command/BitBLT FIFO buffer.

I generally summarise:

Basic accelerators only feature BitBLT & a hardware cursor.
The next best have memory-mapped IO.
The next best have a command/BitBLT buffer, with VRAM being better at higher res/colours. Better drivers.
The next best are 64-bit.

A lot of manuals, especially for the WD90c33, list every minor improvement; buffering, use of off-screen memory, but they're all bested by 64-bit BitBLT'ing.

But what would be a ~2x gain for 64-bit BitBLT'ing is only a 10-20% gain. So the ISA clock becomes a limiting factor.

Those 16-bit memory-bandwidth cards might be the top cards for 286/386sx's. Many drivers for the faster cards only support 32-bit CPUs.

yeah these features do make some difference, but mostly by improving efficiency of existing features, like upgrading a 386 to 486. but having bitblt is something between yes and no, like having a 387 or not.
also i think most isa vga cards with 256kb or 1mb max are 32bit ram width, some like trident8900 would run with 8bit or 16bit but has 32bit for standard configuration. the only exceptions that i know of are: cirrus5401 is 8bit(256kb max), most cards supporting four 4*256kbit dram chips (512kb max) including cirrus5402, trident9000, wd90c11, realtek3105 are 16bit. s3 924 is blamed by some to be 16bit but i can't find datasheet.

MikeSG wrote on 2024-12-19, 05:48:

Those 16-bit memory-bandwidth cards might be the top cards for 286/386sx's. Many drivers for the faster cards only support 32-bit CPUs.

Well, the 386sx is a 32 bit processor. If code runs on a 386dx, it will also run on a 386sx. Programmatically they are identical. From a performance standpoint, the 32 bit memory reads/writes to main memory are slower with the 386sx. But if you are doing memory mapped access to the video card over an ISA bus, they would be the same, yes?

douglar wrote on 2024-12-19, 16:19:

MikeSG wrote on 2024-12-19, 05:48:

Those 16-bit memory-bandwidth cards might be the top cards for 286/386sx's. Many drivers for the faster cards only support 32-bit CPUs.

Well, the 386sx is a 32 bit processor. If code runs on a 386dx, it will also run on a 386sx. Programmatically they are identical. From a performance standpoint, the 32 bit memory reads/writes to main memory are slower with the 386sx. But if you are doing memory mapped access to the video card over an ISA bus, they would be the same, yes?

That sounds right.

Edit: I just tested an S3 928 on a 386sx and it wouldn't boot. It boots on a 386DX. So there must be some check/limit that stops them working.

I can confirm that a S3 928 will run a 386sx system. It might be mainboard related your issue.