[jorvi]

Yeah, no. GDDR is functionally very different than SDRAM.

GDDR tries to push out as much bandwidth as possible, because that really matters for (traditional) GPU workloads. A constant but insignificant (= correctable) error rate is considered completely fine for GDDR, because that sacrifice allows the memory to be pushed much farther.

Meanwhile most (traditional) SDRAM workloads don't give a hoot about bandwidth but really care about latency. And ideally you want no errors, hence ECC RAM being so venerated.

If you unify memory, you're gonna have to choose to sacrifice one of those workloads or go suboptimal for both.

Weirdly enough this mostly matters for non-gaming workloads. The Apple M-series are absolute monsters in gaming, completely crushing the RTX XX90 editions in performance-per-watt, but as soon as memory bandwidth becomes paramount the M-series falls heavily behind.