100% this. David Wang said a few years ago that this was impossible due to the ridiculous bandwidth requirement needed to do chipsets for GPUs.
I still wonder why AMD even tried chiplet GPUs in the first place. They must have thought it would work at first otherwise the 7900xtx wouldn't exist and then suddenly they backtrack and it's gone. I can only assume AMD at first did not think the bandwidth requirements would be as ridiculous as it turned out to be, but also, how could they not know, are these things not able to be emulated in design software?
I do commend AMd for trying new things that Nvidia is too scared to do - AMD put HBM memory on gaming cards, Nvidia won't and AMD was first to chiplets too.
I do hope AMD revisits chiplet GPus again in future once TSMC is able to do fiber optic chiplet interconnects so they can get the bandwidth that's needed to make it work. We're about 5 years away from this. I'll be excited once we get there, the idea of having 2, 3, 4 or more GPUs on a single graphics card is salivating, imagine the performance potential
Last edited: