There are some bus widths that are used in full discrete GPU dies and some that I’ve only ever seen in cut-down GPUs. Full-die GPU memory bus widths seem to always be:
64-bit × 2n × (1 or 0.75)
where n is a whole number. Cut-down models can feature pretty much any multiple of 32 bits.
Examples:
- 128-bit bus: common in entry-level GPUs (RTX 4060, RX 6600 XT)
- 160-bit bus: occasionally shows up in cut-down designs (Arc B570)
- 192-bit bus: common in midrange & entry-level GPUs (RTX 5070, Arc B580)
- 256-bit bus: common in midrange GPUs (RX 9070 XT, RTX 3070 Ti)
- 320-bit bus: occasionally shows up in cut-down designs (RX 7900 XT, RTX 3080)
- 352-bit bus: Appeared in the RTX 2080 Ti, which was cut down
- 384-bit bus: common in upper-midrange & high-end GPUs (RX 7900 XTX, RTX 4090)
Any insights into why this is? As a layperson, it seems like having a full die with perhaps a 160-bit bus for the entry level or a 224-bit bus for the midrange would at least occasionally make sense.
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.
I understand the concept of fusing off or otherwise blocking defective die portions. What I am curious about is the apparent absence of, say, 320-bit bus designs being made which could be cut down as needed to 256-bit.