Anonymous Coward wrote on 2026-04-22, 06:04:
Many of the good 1MB cards with accelerated graphics were older designs that did not support higher colour depths.
Anyone know if the S3 928 takes a big performance hit when only 1MB is installed?
If the S3 928 is used in a straight-forward configuration, it has 32-bit data path from the accelerator to the video RAM for drawing purposes, and a second 32-bit data path from the second port of the VRAM for display purposes. This path can either go directly into a high-performance DAC with 32-bit data input, or it can be handled by the S3 chip and send through a legacy 8-bit data port to the DAC. A card designed that way does not take a performance hit if only 1MB is installed, as the only bank configuration of the 928 is 1MB per 32-bit bank. Adding extra banks does not increase performance.
The S3 928 does offer a special operating mode for high-performance cards (where "high performance" refers to the displaying part, not the drawing part, that is it enables higher resolutions / refresh rates), in which the data path between the RAM and the DAC is 64 bits wide. This can either be used with a DAC that has a 64-bit wide input, or you can multiplex the 64-bit VRAM output into 2*32 bits on the DAC input, allowing higher DAC input frequencies than the VRAM provides as output frequency. S3 calls this operating mode "parallel addressing". For the display data path, the implications of that mode are obvious: You get twice the width and thus twice the bandwidth. As the S3 chip is not involved in transferring pixels from the VRAM to the DAC in these high-performance modes (only for legacy VGA modes), coping with twice the bandwidth is entirely left to the card designer, and the 928 has no bottlenecks that prevent "twice the bandwidth" from being used. Using this mode also means that drawing (which only has a 32-bit data path) needs to be performed accordingly, which essentially uses the same access pattern as bank interleaving. The datasheet does not contain any hints that the S3 chip has increased bandwidth for drawing due to the interleaving nature, but it also explicitly states that it does not gain a drawing speed advantage in that mode. I'd assume the drawing engine is already running at maximum performance without bank interleaving, so the accelerator-side interleaving is just for allowing the 64-bit pixel data path, not for better drawing performance.
The choice between a 32-bit and a 64-bit display path is fixed in hardware. There are cards that use a 64-bit pixel data path, e.g. the miroCrystal 16S, 24S and 32S cards, but you can not find those cards with just 1MB of RAM. So even if the "parallel âddressing" mode does provide performance advantages (which I doubt, as I said), you can not profit of those advantages by upgrading a 32-bit card from 1MB to 2MB.