After establishing that Trident VGA cards are indeed very slow, the natural follow-up question is: Are there cards even slower than that? But not some 8-bit VGA card from the 1980s (or a Realtek from the 1990s), and not some exotic CAD accelerator, but rather some mainstream 16-bit ISA graphics card.
Those who remember the latter days of ISA graphics cards may guess the “winner”: Early S3 graphics accelerators! Such as this one:
This specimen from mid-1992 was not a cheap card, but it was popular. It was a graphics accelerator with 1MB VRAM (not DRAM) and an 80 MHz DAC. The S3 86C911 chip kicked off a very successful line of graphics adapters.
Yet the VGA performance of this card is… well, terrible:
ESCOM VGA 2150D (S3 911): 3DBench : 20.0 FPS Vidspeed VGA : M/W/R 1.1 / 1.1 / 1.0 MB/s Vidspeed MODEX : M/W 1.1 / 1.1 MB/s Doom timedemo : 3820 realtics
The card is clearly operating in 16-bit mode and 8-bit transfers are even slower. The VGA performance is just uniformly awful. These cards weren’t sold for VGA performance but compare poorly in this regard with similar offerings from e.g. ATI (mach8 and mach32 accelerators).
The VGA performance is clearly worse than the horrid Tridents, although the Tridents have no accelerator as a redeeming feature.
A very similar card with updated S3 924 chip is faster, or rather slightly less glacial:
ESCOM VGA 2150D (S3 924): 3DBench : 24.3 FPS Vidspeed VGA : M/W/R 1.4 / 1.4 / 1.2 MB/s Vidspeed MODEX : M/W 1.4 / 1.4 MB/s Doom timedemo : 3227 realtics
The cards do sport an array of jumpers, as usual poorly documented. This appears to be a good match. Unfortunately changing the jumper settings had absolutely no discernible effect.
XGA
So how about the VGA performance of truly exotic cards? How about a Radius MultiView XGA (yes, that’s an ISA card):
This card, other than being an XGA implementation (not a clone; Inmos manufactured the XGA chips under a license from IBM), was notable for supporting multiple accelerators in a single system. Once again, it was not sold for its VGA performance, and it shows:
Radius MultiView XGA (Inmos/IBM XGA-2): 3DBench : 37.0 FPS Vidspeed VGA : M/W/R 2.8 / 2.8 / 1.6 MB/s Vidspeed MODEX : M/W 2.8 / 2.8 MB/s Doom timedemo : 2403 realtics
It’s much better than the contemporary S3 911/924, but noticeably worse than ATI accelerators, leave alone the fastest ISA VGAs.
Cirrus Logic
So are there any other fast cards? Yes, and again they’re Cirrus Logic chips. A card with the CL-GD5402 chip is an AVGA2 with Cirrus Logic branding. Performance is identical to that of an AVGA2 card as expected.
A newer CL-GD5422 based card was also tested. This was a continuation of the AVGA line produced after Acumos had been acquired by Cirrus Logic.
Here are two sets of benchmark results for this card:
Cirrus Logic CL-GD5422 (JP1 open): 3DBench : 50.0 FPS Vidspeed VGA : M/W/R 4.4 / 4.4 / 2.9 MB/s Vidspeed MODEX : M/W 4.4 / 4.4 MB/s Doom timedemo : 1873 realtics Cirrus Logic CL-GD5422 (JP1 closed): 3DBench : 55.5 FPS Vidspeed VGA : M/W/R 5.5 / 5.5 / 2.9 MB/s Vidspeed MODEX : M/W 5.5 / 5.5 MB/s Doom timedemo : 1521 realtics
The JP1 switch is not labeled but presumably selects 0WS (zero wait state) operation when closed, similar to this card. At 0WS, the CL-GD5422 card is slightly faster than the AVGA2 (2.9 MB/s chained VGA reads vs. 2.7 MB/s) and edges slightly ahead of Diamond SpeedStar cards in DOOM performance (but not VGA read speed).
This is a solid card but remember, it depends on the ISA OSC signal and might not work reliably in every motherboard.
Conclusions
So what have we learned? The VGA performance of a card is in no way proportional to its price. But it’s not inversely proportional either—some cheap cards are excellent (AVGA/Cirrus Logic) and others are absolute dogs (Trident, Realtek).
The two families so far identified as having excellent VGA performance are the Tseng ET4000AX and the Acumos AVGA2/Cirrus Logic chips. Both are relatively basic ISA SVGA cards but their VGA performance is hard to beat.
The caveat is that not every card with those chips is fast. In the Acumos/Cirrus Logic case, wait state settings make a big difference. In the ET4000 case, the amount and type of memory on the card makes a difference (cards with more RAM are generally faster).
And finally, some VGA cards are incredibly slow without being “broken”.
Cirrus Logic GD5429 seems to be last in the line of AVGA based ISA adapters. But I replaced here a Spea V7 Vega (GD542x) in favour of ATI Graphics Ultra that has definitely better OS/2 drivers and saves one slot for my Microsoft mouse. Do you know of any OS/2 1.x graphics benchmarking tool? I could only find old PC Magazine’s pmbench.
One weird thing about ET4000AX cards. When installed in my 486, it actually causes the AMIBIOS POST memory check to run noticeably faster. It doesn’t do it with any other video card I have, including VLB cards.
Another “slow poke” VGA chip to test would be the Oak OTI-087. Diamond is famous for putting this stinker on the Viper VLB card leading to extremely poor performance with DOS games. Later revisions of the board came with a Weitek 5186 on the VGA side of things, no clue how it compared to the Oak chip performance wise.
Not aware of any surviving benchmarks… you’d probably have to roll your own, just run a bunch of Gpi calls that can/should use hardware acceleration (BitBlts, rectangle fills, line draws, text output).
I unfortunately don’t have any OAKs at hand at the moment. From what I remember, they were neither very fast nor very slow, but I could be wrong. I don’t think the Weitek 5186 was all that great either.
The thing about POST memory check could be optimized text BIOS, maybe also fast VGA text mode. In some implementations, the memory test is a text output benchmark.
A bit of a necropost, but despite its ill-starred appearance on Weitek P9000-based video cards, the OTI087 is most definitely *not* a slow chip. The reason why it performed so poorly on those cards is because it was usually given an 8-bit data path to its dedicated VGA memory (which was separate from the P9000’s memory), so it couldn’t help but be sub-par.
The OTI087 is frequently cited as an ‘accelerator’, but the reality is, it’s just a near-equivalent to the TC6059 ET4000AX performance wise, with great host interface throughput for an Oak chip, but no actual acceleration hardware beyond a hardware cursor and color expansion registers. Like the ET4000AX, its performance varies with the memory configuration and card design.
One question … what benchmark software did you use ?
I could add some numbers from my video cards, like I’m wondering if it is any improvement between the different ISA Trident chips flavors ?
Back in the days (was assembling/selling computers in the early 90’es) I never really bothered to do a speed test under DOS. But one thing is sure: Trident was terrible under Windows. With the SVGA drivers installed (256 colors, 800×600) you could actually see how the windows are drawn on the screen 🙂
And then came the first (buyable with decent money) alternative: Cirrus Logic
And suddenly the windows were flying. It was one of these “Ahaaa” moments
The benchmark software is listed there — it’s 3DBENCH.EXE, VIDSPEED.EXE, and DOOM (I think I used DOOM 1.1 shareware for the tests).
The Trident chips were slow, and the cards which used Trident chips were even slower (similar to OAK chips). Basically the chips weren’t speed demons to begin with, but they were put on cards with slow RAM and 8-bit bus and/or memory interface, and that would kill the performance of any chip.
Cirrus Logic was indeed unique in that they made graphics chips which were generally very solid, reasonably fast, yet cheap.
Is 8 years too late to make a comment?
The top photo shows the VGA card that uses a Music Semiconductor VGA Palette chip. This chip was a drop-in replacement for the original Inmos chip used by IBM in the original VGA board. There were several improvements, such as a much cleaner DAC design. The Inmos chip had a significant amount of glitch energy when the DAC output changed from 011111 t0 100000 (or the other way around) as the worst case. The Music chip improved on this. Showing vertical bars on the screen shows the difference.
Another improvement was that the Inmos chip could only support a change to the palette RAM during retrace, otherwise the monitor would see annoying snowflakes. The Music chip allowed the palette RAM contents to be changed at anytime, allowing animation and other software features to use the Palette RAM.
No, there’s no “too late to make a comment” if it’s relevant. And it is, thanks!