If I remember correctly, their cards had a high compute rate even at double precision, but had tiny memory buffers and basically couldn’t do texturing at all in practice.
It turned out that double precision was a mistake that was sold as a “professional” feature. By sharing edges correctly and using the correct rounding modes, single precision provides pixel-perfect rendering. Efficiencies like this allowed the consumer GPUs to run circles around SGI hardware.
Most SGI graphics cards had no texture ram at all, and had zero texturing capability. At the time there was one card that had 4 MB of texture RAM, but in the same year a typical PC GPU had between 16 or 32 MB of shared memory, most of which was used for textures.
A “budget” SGI card cost more than my parents’ car. I bought three different GPUs by that point with my pocket money.
I honestly don'tt know what you're talking about- I wrote many programs that did high performance texture mapping on SGIs and they had both texturing capability and RAM. When you say "SGI card", it makes me sound like you're talkijng about something other than an Octane2 or InfiniteReality.
The Indy/Indigo2 graphics - the Elan/Extreme/XZ - never had texture hardware, but would fall back to software. It wasn't until the High IMPACT series was released for Indigo2 in 1995 they had hardware texture mapping on their "low end" - i.e. desktop sized systems.
1995 is when I started using SGIs in earnest (on Indigo2 and RealityEngine). I do recall a lab next to ours that had $15K indys that were absolutely useful for doing anything with graphics.
You’re very confused. In 1992 you could get an SGI system equipped with 16MB of just texture RAM. Long before 3D accelerators for PCs. Later iterations such as InfiniteReality of course had much more texture memory. The SGI VPro could be configured with 108MB for texture memory.
It turned out that double precision was a mistake that was sold as a “professional” feature. By sharing edges correctly and using the correct rounding modes, single precision provides pixel-perfect rendering. Efficiencies like this allowed the consumer GPUs to run circles around SGI hardware.