Hardware

Part of this has to do with inevitable flaws in manufacturing. Nothing is ever totally perfect.

GPUs have many multiple identical cores running in parallel to achieve the processing power they do. But during manufacturing, let's say for simplicity that the GPU is designed with 64 cores. But in testing, only 57 of those cores actually function as expected.

So what are they gonna do? Well they're damn sure not going to scrap the entire chip, they just basically block out the bad cores and sell it as a lower end model with reduced processing power. I'm practically certain this also affects the potential bus width of the final product.

It’s related to the architecture and the number of chips used. Repeating or standardizing on a lower bus size for your low-med range makes it simpler, cheaper, faster to manufacture

Post in thread 'Explain to me how memory width (128/192/256 bit, etc) is related to memory amount.' https://www.techpowerup.com/forums/threads/explain-to-me-how-memory-width-128-192-256-bit-etc-is-related-to-memory-amount.170588/post-2699537

If I'm not mistaken, most memory chips use 32-bit-wide interfaces, meaning you are stuck with multiples of that. They come in multiple capacities, I think around 1 or 2 GB ones (including 1920 MB / 2 GiB / whatever) are most common these days. The 2 GB ones are higher density, but more expensive, so you need to decide what you care about first. You also don't want to let to bandwidth go too low, so your performance doesn't tank.

You then decide much capacity you want on your card. That decision is primarily driven by cost and demand, but you also don't want the gaps between SKUs to be too small. It just doesn't make sense to have tons of SKUs very close together, neither for you, nor your competitors. So the industry came up with a difference of 2 chips ≈ 2 or 4 GB as the minimum difference.

We see 4, 8 12, 16 and so on GB cards, but rarely something like 5 or 10, because for 8+ GB, you would typically use the 2 GB chips, so you have 4 GB jumps there, and for the sub-8-GB ones you have 1 GB chips, hence 2 GB jumps. This leads to 64 bit jumps in bus width, so 128, 192, 256, 320... bit bus-widths.

128 is already very low in terms of bandwidth, I don't know if you can actually find this on any modern cards, but that's about the reasoning here.

And I guess we don't start at 3 or 5 GB because it's just weird, I don't know :D