ecmaster76
[H]ard|Gawd
- Joined
- Feb 6, 2007
- Messages
- 1,150
I wouldn't say their engineering didn't screw up. 2 years after launching the Fury and they managed to ship a card that has less memory bandwidth than its predecessor while using HBM2.
HBM2 didn't meet its design targets and that seems to be Hynix's issue. AMD gambled that 2 stacks would be enough (and maybe it isn't the bottleneck on Vega anyways). From an assembly perspective it has to much, much cheaper to attach 3 parts to a smaller interposer in Vega than the 5 pieces and huge interposer Fiji required.
Supposedly the Fiji interposer "die" was over the reticle limit for the fab process it was on which gets expensive in a hurry. If Vega is under that then they made the right choice there
Almost certainly all the shortfalls of the performance are architectural though I could believe HBM2 contributed to Vega being so late (and really that's the bigger issue)