I would rate top-to-bottom:
(Edited this many times to make a more formal list. Grouping describes general performance)

(3D)
1: Artist Graphics 3GA (VRAM 64-bit - 3D gouraud shading, 3D wireframe, 3D dithering, 3D texture mapping, 6x 3D raster ops, Z buffer, 2D accel & VGA - AVGA core). OpenGL support.
1: Artist Graphics GPX (VRAM 64-bit - 3D gouraud shading, 3D wireframe, Z buffer, 2D accel, separate DOS VGA chip - CT450). 3D in CAD only. *
1: Matrox Impression Plus 220 (VRAM 64-bit - 3D gouraud shading, 3D wireframe, Z buffer, 2D accel). Video For Windows 1.1D playback acceleration.

(2D - 64-bit, high res)
1: ATI Mach 64 (VRAM 64-bit - Packed 24bit drawing, advanced drawing modes, command context chaining).
1: IIT AGX-016 (DRAM/VRAM 64-bit) **
1: Oak Spitfire OTI-64105/107/111 (DRAM FP/EDO 64-bit) ***
1: CL GD5434 (DRAM 64-bit)
1: Sierra Falcon64 (DRAM EDO 64-bit) ****
1: ET 4000 W32i (DRAM interleaved 32-bit)

(2D - 32-bit, high res)
2: S3 928 (VRAM 32-bit, offscreen buffer, command FIFO, pipelining). Specialises in high res/high colour. *****
2: ATI Mach 32 (VRAM 64-bit, 32-bit 2D accel, offscreen buffer, command FIFO). Specialises in high res/high colour.
2: IIT AGX-015 (VRAM 32-bit). Specialises in high res/high colour. Specialises in high res/high colour. **

(2D - 32-bit, fast)
3: Chips F65545 (DRAM 32-bit, large offscreen frame buffer, command FIFO, optimised drivers). Specialises in 640 x 480 x 8-bit. ******
3: WD90C33 (DRAM 32-bit, offscreen buffer(?), command FIFO) *******
3: Oak OTI-087 (DRAM 32-bit, up to 2MB, 80Mhz, command FIFO/cache)
3: Weitek 5286 (DRAM or VRAM 32-bit - offscreen buffer, command FIFO) ********
3: TGUI9440 (DRAM FP & EDO 32-bit, command FIFO)

4: S3 924 (VRAM 16-bit)
4: IIT AGX-014 (VRAM 16-bit)
4: CL GD5429, WD90C31 (memory-mapped IO)

5: Other S3, CL, WD, TGUI

This is flawed in multiple ways... They should be tested on the ISA bus at the same clock divider, in WinTach/WinBench and in Starcraft.

Updated this a few times to include more cards. Everything in category 2 is basically 1993+ with more than memory-mapped IO. All groups do hardware BitBLT.

The Artist Graphics 3GA is considered the fastest due to magazine/marketing rumours of it being faster than an S3 968 (both in PCI). S3 made the fastest 2D PCI cards except for Matrox.

* Artist GPX (ISA) is used in the Artist Graphics card (P/N 90070-7025)
** IIT AGX-015/16 is used in the Hercules Graphite Pro/Power, Orchid Celcius
*** Oak Spitfire OTI-64105/107 datasheet specifies 16-bit ISA compatibility but unknown if physical cards exist.
**** Sierra Falcon64 is used in the Techworks Thunderbolt64, STB Horizon 64
***** S3 928 is used in the Number Nine (#9) GXE, Spea V7 Mercury
****** Chips F65545 is used in the Advantech PCA-6653
******* WD90C33 is used in the WD Paradise Accelerator 24
******** Weitek 5286 is used in the Alpha Systems Lab Inc

Mach32 has 64-bit memory.
Does it really rank below et4000w32p?

In the windows spreadsheet posted on the previous page, the ET4000 W32i was ~10% faster than the Mach32.

Using WinBench 3.11 256c scores. Calibrated by the S3 Trio32 2MB scores.

Benchmarked a Chips & Tech 65545 and Cirrus Logic GD5429 ISA card/s today.

The Chips & Tech is astoundingly fast in 2D for a 32-bit card....

WinTach 1.0.

For comparison, from previous page (Cyrix 5x86 @ 100Mhz, 3x33Mhz, 11MHZ ISA Clock):
ATI Mach64 2MB ISA:
640x480x8 (256 colours) 32.35
640x480x16 (32k colours) 72.4

CL-5434 2MB ISA:
640x480x8 (256 colours) 35.35
640x480x16 (32k colours) 63.25

~

My tests. (Intel DX4-100, 33x3MHz, 16MHZ ISA Clock)
CL-G5429:
640x480x8 (256 colours) 23.04
640x480x16 (32k colours) 40.60

Chips & Tech 65545 ISA, win95c drivers
640x480x8 (256 colours) 37.51
640x480x16 (32k colours) 59.18

~

The Chips & Tech card is faster than both Mach64 & CL-GD5434 in 8bit mode. Slightly slower in 16bit.

Ran the tests twice.

The CAD score (primarily full screen scrolling) is ~3x faster than the CL-GD5429. Paint is much faster too.

Full screen scrolling in Starcraft is very responsive. The 5429 is more like a slideshow. Units move fast enough on both.

Highly recommend for a 386->486 project, except some of the RAM seems to be used for buffering. It's 1MB but only supports 800x600 in 256 colours max, or 640x480 in 32k colours.

Chips 65545 (vs CL-GD5429)
(8-bit)
Word Proc: 16.18 (even)
CAD: 76.52 (+292%)
Spreadsheet: 27.64 (even)
Paint: 29.68 (+170%)

(16-bit)
Word Proc: 24.24 (+20%)
CAD: 135.00 (+260%)
Spreadsheet: 29.12 (even)
Paint: 47.32 (+46%)

MikeSG wrote on 2024-10-04, 13:39:

I would rate top-to-bottom: ATI Mach 64 (64-bit) CL GD5434 (64-bit) ET 4000 W32i (interleaved 32-bit) ATI Mach 32 Chips F65545 ( […]
Show full quote

I would rate top-to-bottom:
ATI Mach 64 (64-bit)
CL GD5434 (64-bit)
ET 4000 W32i (interleaved 32-bit)
ATI Mach 32
Chips F65545 (drivers claim to improve performance by +80%)
S3 928, CL GD5429

RAM & drivers matter a lot in windows acceleration

hello, i see you placing the cirrus5429 on the bottom. i am not challenging that but just think, cirrus still had the older and even slower 5428 and 5426, could the 5426 be the slowest graphics accelerator in the world?
also, the cirrus5430 is probably the slowest pci video card aswell.
the chips 65545 seems to perform very well, but supporting only 1mb of ram is a major drawback.

That's the bottom of the top, not the bottom of the bottom, there's a lot below that. 5426 and 5428 are generally similar in performance, given fast RAM and clock defaults on the implementation to the 5429, they just gain minor improvements and features. The 5424 and 5422 are older members which are not as good. Then there's a other families of VGA chipsets, TVGA, OTI, WDC, UMC, Realtek...

BitWrangler wrote on 2024-10-27, 17:57:

That's the bottom of the top, not the bottom of the bottom, there's a lot below that. 5426 and 5428 are generally similar in performance, given fast RAM and clock defaults on the implementation to the 5429, they just gain minor improvements and features. The 5424 and 5422 are older members which are not as good. Then there's a other families of VGA chipsets, TVGA, OTI, WDC, UMC, Realtek...

yeah but this topic is about isa vga cards with graphics acceleration, not all isa video cards. the tseng et4000ax, cirrus5422/5424, trident8900d/dr and wd90c30 are all plain cards without acceleration, yet they had all hit the limit of isa bus and therefore should all perform identically. i doubt if trident, oak, umc and realtek had ever made any isa cards with acceleration.
in all, cirrus was a name i favored in the 2d years. its performance may not be the best but surely decent and affordable, and quite compatible aswell. there was no major drawback i could name save for the lack of ega fonts, until it failed miserably in the 3d era, only nvidia and ati made it out.

Updated the post ranking... They're all close in performance for general windows accelerations. There's a small step to the 5429/5428, and a small step to other accelerator cards from what I've seen.

After seeing the ET4000 W32i (ISA) run Starcraft so smoothly in this video, it'd be interesting to rank cards by Starcraft performance... https://www.youtube.com/watch?v=Q-47q33atbk
The Chips 65545 card I tested was not as fast as this (5-20fps), but was miles ahead of the 5429 (1-10 fps).

Diablo is about 5fps too slow to be playable.

There's a whole bunch of S3 928's with a 32-bit RAMDAC on Ebay at the moment if anyone's interested to test them.

They are 2MB 32-bit VRAM. ISA.

Number Nine GXE (S3 928) - 1992. 3 avail. https://www.ebay.com.au/itm/186703717638
Number Nine GXE (S3 928) - 1993. 14 avail. https://www.ebay.com.au/itm/186700521610

As far as Windows 3.1 on a 286 goes, here's some results on my Headland HT18 mobo with a Harris CPU running at 24Mhz and the ISA bus running at 12Mhz - I should note that the CPU performance is more or less in the middle of a HT12 with 0WS or 1WS, and there's no way to configure wait states in the BIOS. Unfortunately, a 25Mhz 0WS setup with my HT12 mobo is unstable though I have seen better results on that board, this will do for now:

1024x768x256

(apologies for 88% resolution, currently using my X230 Thinkpad which does not have a 1080p LCD upgrade)

I added the GD5420 in the screenshot because the GD5429 can only use GD5422 drivers (only v1.10 and earlier, everything else requires a 386 and the earliest GD5426 drivers I've seen are v1.20) but even then there's a notable performance increase plus it has better font rendering speed compared to the ALG2101, the ET4000/W32i and the WD90C31. I also added the ET4000AX to show that the ET4000/W32i doesn't seem to do any significant acceleration with its drivers in Standard Mode, but it does appear that interleaving is working. The mach32 cards use the mach8 drivers - they have glitchy output in the final paint scene in WinTach, hence why they're marked. The DRAM model would not let the co-processor work at 16-bit with my HT18 mobo, so only 8-bit results are shown - I haven't tested it with my HT12 mobo yet.

The ALG2101 is unique in that it loses performance when going higher in resolution. The only useful results (apart from Winbench, which I don't trust) are from Windsock as the benchmarks run in the same sized window across all resolutions:

640x480x256:
Word Processing 117
Spreadsheets 155
CAD 108
Paint/Draw 155
Overall 127

800x600x256:
Word Processing 111
Spreadsheets 147
CAD 105
Paint/Draw 147
Overall 121

1024x768x256:
Word Processing 105
Spreadsheets 138
CAD 101
Paint/Draw 138
Overall 114

Since I made those earlier posts in the linked post, I have learned that the S3 801 does work fine in my HT12 mobo with the S3 911/924 v1.7 drivers, and that the performance is an improvement over the 924. I also picked up a Diamond Stealth 24 with a 924, which also exhibits the same unstable behaviour on my HT18 board but is perfectly fine on the HT12 board. For some reason, the STB Powergraph X-24 has gimped DOS performance compared to another 801 card that I had (Photon Torpedo ISA) which had unfortunately died too quickly. The ET4000/W32i's greatest strength is that it has a very high Speedy score in 16-bit compared to most other cards. The GD5434 and S3 928 require a 386 to POST - even then, the S3 928 is not compatible with S3 drivers that work on a 286.

The biggest anomaly however is the Weitek 5286 - the existing drivers are all for the 5186 and the 256-colour modes are completely useless while the 16-bit modes don't appear to accelerate, but the 16-colour modes tell a different story. Unfortunately for some reason the 1024x768x16 mode crashes to DOS while keeping the framebuffer at 1024x768, but 800x600x16 shows an interesting result (Cirrus Logic drivers I tested were having some specific issue with this mode that I don't recall):

If there is one useful thing about Winbench, it does show separate benchmark results for horizontal, vertical and diagonal lines - the WD90c31 does not accelerate diagonal lines at all, while the ALG2101 has lower vertical and diagonal line scores compared to the ATI and S3 cards. The WD90c31 also has extremely suspicious blitter results in Winbench, somehow getting 50% higher scores compared to the ATI and S3 cards.

Finally, I'm not sure how to categorise the OTI087 and OTI087X - they're significantly better than the ET4000AX in just about every way, but they also don't compare to other accelerators (apart from obvious weak points in the ALG2101 and WD90C31, the OTI087 does compare somewhat favourably to the GD5249 on a 286 though the GD5249 still has better font rendering) plus the OTI087X has a bad habit of frequently being gimped on memory bus width (I've had a card with four 512K x4 SOJ-20 chips for ages, then finally got one with six 256K x4 DIP-20 chips, two 256K x3 (?!?!) DIP-20 chips and two 256K x1 SOJ-18 chips last week) and having both performance improvements and regressions compared to the OTI087, at least with a quick test on a 386SX-40.

MikeSG wrote on 2024-10-27, 09:15:

Benchmarked a Chips & Tech 65545 and Cirrus Logic GD5429 ISA card/s today.

Have you had any success with Win 3.1 drivers for the 65545? Not sure if there's something funny with the BIOS on my card (I know that 655xx chipsets have a wide range of different BIOSes for video output - can you please provide a dump of yours?), but I get garbage output and the system crashes regardless of what system I throw it in. 65535 drivers do work, but they're unaccelerated.

MikeSG wrote on 2024-11-09, 11:35:

There's a whole bunch of S3 928's with a 32-bit RAMDAC on Ebay at the moment if anyone's interested to test them.

They are 2MB 32-bit VRAM. ISA.

Oof, I just bought a NIB 2MB Diamond Stealth Pro from the US for slightly more than these GXEs, but there's enough hardware differences between that and these GXEs (looks like the second link is 3MB, which is consistent with the maximum capacity of their PCI 928 cards) that I decided to pull the trigger again. Will find out in a few weeks when they arrive.

The S3 cards are interesting in your tests.. I can't work out the difference between them and similar cards except for running slightly faster clocks.

DEAT wrote on 2024-11-11, 05:07:

[...]the WD90c31 does not accelerate diagonal lines at all, [...].

If you're interested in testing a WD90C33, the model after, the manual says they are compatible with ISA. I don't think any actual ISA cards were made only VLB/PCI, so it would mean transferring the chip from one card to another, assuming pins are compatible. Drivers might be an issue though.

DEAT wrote on 2024-11-11, 05:07:

Have you had any success with Win 3.1 drivers for the 65545?

I haven't tested Win3.1, only Win95c.
My card is an AdvanTech PCA6653. The drivers are still available on the website (Win3.1, WinNT, Win95, others): https://www.advantech.com/en-au/support/detai … ver?id=1-18MWVQ
In Win95, the driver to use is a generic name like "accelerated vga" or something like that. Not one of the list 65545 drivers they're all PCI.
I could only get 640x480 to work. Other resolutions pan the screen around when you move the mouse.

I found this recently... A man by the name of Boogeyman wrote a video card FAQ in 1995 reviewing all VGA cards at the time, talking about performance of 2D acceleration, Dos. I don't know how he got access to all of them but he goes into some depth. Describes cards as "okay", "bad", "very good", "superb". Tested on a 486 DX-33, DX2-66.
https://www.chiark.greenend.org.uk/~andrewm/D … eo-card_faq.txt

You can see some benchmarks of a 16bit pcmcia S3 Trio64V+ if you follow the third link in my signature.
I'm curious to know how it compares to all of your cards but I didn't use the same benchmarks as you did in this thread...and I don't have the card anymore to run more tests.

1Speedy speedmarks
2C&T65530 VGA driver 5.90
3C&T65530  CT LINEAR4 driver 7.38
4
5T64V	  640x480 256 19.13
6		640x480 32k 13.76
7		640x480 64k 13.76
8		640x480 16M 8.56
9		800x600 256 21.60
10		1024x768 256 23.76
11		1024x768 32K 16.58
12		1024x768 64k 16.63 

You can find Speedy (if you want it) at https://archive.org/download/So_Much_for_Wind … LS%2FSPEEDY.ZIP

Thanks all for submitting the results. Very interesting. For Mikes link: good details but the benchmarks seem to be done with fully different systems per videocard. This makes comparison difficult.

Comparing the CL GD-5429, C&T 65545, and S3 928 cards in WinTach 1.0.

DX4-100, 33x3Mhz, 16MHz ISA clock. win95c drivers.

WinTach 1.0

CL-G5429 (Jaton):
-640x480x8 (256 colours) 23.04
-640x480x16 (32k colours) 40.60
Chips & Tech 65545 (Advantech PCA-6653)
-640x480x8 (256 colours) 37.51*
-640x480x16 (32k colours) 59.18
S3 928 (NN GXe):
-640x480x8 (256 colours) 28.96
-640x480x16 (32k colours) 66.91*

Comparison vs slowest, highest score shown:

8-bit:
-Word Proc: 16.18 (even)
-CAD: 76.52 (+292% C&T 65545, +184% S3 928)
-Spreadsheet: 27.64 (even)
-Paint: 29.68 (+170% C&T 65545)
16-bit:
-Word Proc: 38.04 (+190% S3 928, +20% C&T 65545)
-CAD: 135.00 (+260% S3 928, C&T 65545)
-Spreadsheet: 38.72 (+30% S3 928)
-Paint: 60.16 (+185% S3 928, +46% C&T 65545)

The Chips & Tech 65545 was faster in 8bit mode, but the s3 928 was much faster in higher colours/higher res. The NN GXe can do 1280x1024x8, and 1024x768x16.

The S3 928 seemed as fast as the Chips & Tech 65545 in Starcraft. I don't think any of the 32-bit cards are fast enough for Starcraft except for the ET4000 W32i.

thanks for that benchmark.

MikeSG wrote on 2024-11-11, 07:45:

I haven't tested Win3.1, only Win95c.
My card is an AdvanTech PCA6653. The drivers are still available on the website (Win3.1, WinNT, Win95, others): https://www.advantech.com/en-au/support/detai … ver?id=1-18MWVQ

I decided to pick up a PCA6653 from eBay and was having the same issue with my existing F65545 card - eventually did some deeper research and it turns out that having >=16MB of system RAM makes the Win 3.1 driver blow up. Random thought process right at this minute is that there's a reason the Memory Hole BIOS option exists, and it's for chipsets like this.

On the plus side, while my existing card has 30ns EDO RAM, the PCA6653 doesn't have VGA palette cycling issues that was plaguing the other card - Wolf3D intro works perfectly fine.

MikeSG wrote on 2024-11-11, 07:45:

I could only get 640x480 to work. Other resolutions pan the screen around when you move the mouse.

You want to use CT.COM, which will force CRT mode - the BIOS is set to "simultaneous display" mode which will display the same resolution on both LCD and CRT output. Easy enough to find, several shovelware CDs have it available:
http://discmaster.textfiles.com/view/7312/199 … s655.zip/CT.COM

The real problem I'm facing is changing the BIOS - C&T cards have a "BMP editor" utility that will modify the BIOS settings and then you can flash the modified BIOS, but I've only found the BMP editor for the PCI F65545 and F65548, along with F6555X cards. There's screenshots on a single blog for a F65535 BMP editor, but I haven't found any specifically for the ISA F65545. Trying to use the PCI F65545 BMP editor spews out an error upon loading the BIOS ROM, and then has invalid BIOS settings in several locations, while trying to change them shows blatantly incorrect values.

Ahh thanks. Was wondering about the screen panning. I did notice in Win95c there's an extra tab next to Display options for Chips & Tech, which is for selecting CRT/LCD. I just left it as "both"... Will try CRT only next time.

Found this PDF on explaining the BMP editor program for F65540/F65545 cards. Chapter 5: https://files.embeddedts.com/ftp/ts-x86-sbc/o … nents/65545.pdf

Says once the BIOS is changed from the original file it cannot be modified again... There is one 32Kb BIOS in the Advantech PCA6653 drivers zip.... Maybe that's an original...

I don't know if that's the problem you're having. Would be nice to change the clock speed.

MikeSG wrote on 2024-11-11, 07:45:

If you're interested in testing a WD90C33, the model after, the manual says they are compatible with ISA. I don't think any actual ISA cards were made only VLB/PCI, so it would mean transferring the chip from one card to another, assuming pins are compatible. Drivers might be an issue though.

I saw an ISA card with WD90C33 pop up on eBay last year. They're rare, but do exist.

I did see that recently... They're in the WD Paradise Accelerator 24.

is it safe to say that the slowest class with acceleration are cirrus5426, oak087 and s3 911?
also i doubt if any isa version of trident9200/9400/9440 exists.