This GPU has a 192-bit memory bus. At 32-bit GDDR bus width (well w/ 6, 2x16b data channels per chip), that means you have 6 channels. With regular GDDR6, the l

This GPU has a 192-bit memory bus. At 32-bit GDDR bus width (well w/ 6, 2x16b data channels per chip), that means you have 6 channels. With regular GDDR6, the largest produced size is 16Gb (2GB), so 12GB is what you get. You could double that up w/ a beefed up PCB if the memory controller supports it to get up to 24GB (in the way workstation cards like W7900 and A6000). Beyond that, you'd have to move to GDDR7 (which has 24Gb/3GB chips incoming) or to HBM stacks, but at that point you're well beyond a "basic GPU". I think the only way you could get to 128GB would be either using regular (LP)DDR or HBM. Note, Apple M chips have weak GPUs with decent MBW and large memory capacities (up to 192GB @ 800GB/s for an M2 Ultra, launched mid 2023) and have not been a major CUDA threat so I don't think your hypothesis actually stands up.