In the late 1980s and the early 1990s, the so-called Bus Wars raged. A few years after the PC/AT was released, it became clear that the ISA bus could not keep pace with faster CPUs and peripherals, especially graphics cards and SCSI HBAs.
IBM’s solution, Microchannel or MCA (1987) was technically excellent, radical, and ultimately a failure. Compaq and the “gang of nine” bet on EISA (1989), a less ambitious but in many ways very similar bus with one major difference—backwards compatibility. EISA was in practice even less successful than MCA and practically unheard of outside of servers and some workstations.
Around 1992, the VESA committee standardized the VESA Local Bus or VLB, geared towards but by no means limited to graphics cards, and designed primarily for the 486’s local bus. For about two years, VLB was very successful and graphics cards designed for VLB were unbeatable.
Around the same time, Intel worked on PCI, a bus which successfully learned from past mistakes. In late 1993, the first PCI systems and adapters became available, and PCI in combination with Pentium systems very quickly destroyed all competition.
For a short while, all five buses (ISA, MCA, EISA, VLB, PCI) existed in the market. Some adapters were available in two or three bus variants and a precious few went all the way. One of those was ATI’s mach32 graphics chip and the adapters based on it, Graphics Ultra Pro and Graphics Ultra + (using VRAM vs. DRAM, respectively).
The mach32-based ATI Graphics Ultra is a true child of the bus wars. In 1994, ATI sold the Graphics Ultra in all five bus variants, something not imaginable any any other point in PC history. The OS/2 Museum presents the entire family: Five ATI mach32 graphics cards built in 1994.
The mach32 was ATI’s first chip with PCI support, the last with MCA support, and the only with EISA support. The MCA, EISA, and VLB units happen to be VRAM-based, the ISA and PCI variants use DRAM. The mach32 cards used a variety of DACs with pixel clocks between 80 and 135 MHz. The maximum resolution was 1280×1024. The VRAM-based cards tended to use better DACs and the image quality was generally very good.
For comparison, a look at the five different bus connectors:
At least the ISA variant was available with a mouse port and came with a mouse. The EISA and VLB PCBs had space for the mouse connector as well (on MCA it made no sense and on PCI it wasn’t technically possible).
It is difficult to make direct performance comparisons because there were no five-bus motherboards and four-bus boards are exceedingly rare. The OS/2 Museum can only run head to head comparisons in PCI/EISA/ISA or PCI/VLB/ISA systems. However, benchmarks are something for a follow-up post.
The ISA-based mach32 boards (such as the Graphics Ultra+ definitely suffered from a bus bottleneck. How much that was felt depended on the application—as long as the accelerator was busy, any differences were negligible. But larger bus transfers were slow on ISA.
The VLB-based mach32 boards might be the most widespread. The bus was noticeably faster than ISA. A VRAM-based Graphics Ultra Pro with a good DAC was a very nice VLB card, with the caveat that the mach32 reportedly did not work with bus speeds beyond 33 MHz.
The PCI variant of the mach32 was a latecomer and required the ‘AX’ variant of the chip. The PCI based mach32 boards tend to have lower speed DACs and DRAM. The newer mach64 chip (no longer 8514/A compatible) was designed for PCI from the beginning and quickly replaced the mach32.
The MCA and EISA variants are quite uncommon, and it’s hard to say which one of them is rarer. EISA graphics cards are generally very scarce and only a handful of models was ever produced. One of the major reasons was the fact that most EISA systems were servers which did not need fast graphics and an ISA VGA card was sufficient.
Note the unused space for a bus mouse on the EISA PCB (lower right corner).
While MCA graphics cards were sold in much larger quantities than EISA, in the early to mid-1990s there were very few non-IBM offerings. The Microchannel market was generally uninteresting to OEMs because the vast majority of sales happened in the ISA, VLB, and later PCI form factors.
ATI’s mach32 was one of the few 3rd party Microchannel graphics cards available in 1994. It is possible that ATI only offered a MCA bus variant of the mach32 because it was a successor to the earlier mach8 chip—which was designed in the late 1980s as an improved 8514/A clone and supported nothing but Microchannel and ISA buses.
It is a good thing that by 1995, the bus wars were effectively over and PC users did not have to choose from five different buses when shopping for a graphics card (although AGP arrived just a few years later to make things more interesting again specifically for graphics cards).
One reason for the rarity of EISA video cards is because many of the motherboard makers quickly added VLBus to their EISA workstation boards. I could do an ISA/EISA/VLB comparison on the same motherboard, but I don’t have any mach32 cards here.
Maybe you have some S3 928 or so ISA/EISA/VLB boards? That would be interesting too. I think EISA/VLB boards are really rare.
You’re probably right that on the high end, VLB overtook EISA very quickly, and on the low end people just used ISA cards. The demand for high performance EISA graphics cards could not have been significant at any point.
My PC plug-in card collection is actually pretty small these days, the only S3 card I have is VLBus. The EISA motherboards came late in collecting, so I only have EISA Ethernet and SCSI cards (I happen to have one of those for every bus for some reason). I might eventually let go of one of the EISA/VL motherboards I got, I really don’t need three (they aren’t rare around here, things sometimes just appear in groups). Much to my surprise two of the boards were Made in USA!
I think most EISA and MCA high-end video cards were sold as part of prepackaged CAD systems so very few retail offerings. The EISA ID list shows about 40 distinct EISA bus video and graphic cards but many are from Compaq, HP, Dell, and DEC. Rarest of the bunch would be the DEC EISA video cards for Alpha AXP produced in the year before DEC moved Alpha to PCI.
Actually Made in USA is not at all surprising for high-end 1990s boards. Do the EISA/VL boards have OPTi chipsets or something else? I’d probably be interested whatever it is… if you need to make more space 🙂
SCSI would be another bus benchmark tool, except Fast SCSI won’t push most of them and I don’t know how many Ultra/Wide HBAs were available for non-PCI platforms. Haven’t done any research though.
Based on anecdotal evidence (boards I’ve personally seen) I suspect the late 1980s workstations used MCA and the early 1990s machines switched to EISA. But yes, the more interesting boards weren’t really sold in retail stores.
I think most non-IBM workstations used ISA before EISA. Non-IBM systems using MCA was never common in the first place (I think Tandy and NCR were a few of the major ones).
Shoot me an e-mail. The one board I’m likely to let go of is a Micronics EISA-VL2. Its a full AT sized Socket 2 486 board that uses a partial implementation of Intel’s 82350DT, the memory interface is custom Micronics instead of the i82359 DRAM controller. It is a full 32-bit bus-mastering EISA board, not a partial implementation like the infamous HiNT Caesar.
The other two boards I have are both SiS 85c411/85c406 based.
My EISA SCSI card is an Adaptec AHA-2740W, so I have a non-PCI wide controller to test with. Sadly its VLB cousin, the AHA-2840A doesn’t have wide bus support.
Thinking about it, I wonder if the proprietary local bus solutions were part of why VLB was rushed so quickly in 1992 even though PCI was coming in 1993 with the Pentium.
BTW the Mach32 ASIC comes in two varieties: 688003 and 688006. You can find VRAM based VLB cards with either and the latter is a great deal faster. So far I’ve seen ISA and EISA boards only with the ’03.
The proprietary local bus (and then VLB) solutions were a cheap hack to improve Windows performance for the masses. TIGA (the previous attempt at providing PC graphics performance) was too expensive, and couldn’t do much to help memory-to-memory copy, which was what Windows 3 primarily needed. EISA, as a proper bus, was also expensive, and not designed for interactive graphics performance. VLB, despite its overt bletcherousness, was a clear win for that.
VLB was a quick hack at the right time. I don’t know the history of pre-VLB local bus graphics, but I know that the C&T Wingine was a thing and that it pre-dated VLB. I don’t know if there was much before 1992.
The thing is that as unreliable and hacky VLB was, the bandwidth between the CPU and adapter cards could not be beat. And unlike MCA/EISA/PCI, there was no extra stuff for software for worry about (like the ability of software to discover and configure hardware).