From the earliest review we have seen that the Fury X card has seesawing performances in video games: sometimes it performs very well, sometimes it performs awful compared to the “old” GPU Hawaii/Grenada. Why?
If we consider the drivers, we know from long time (CeBIT 2015) that AMD has some trouble to write a correct version of Catalyst for Fiji: the HBM management is not perfect, yet. AMD is working hard on a huge Catalyst improvement (Omega 2?), expected in October/November, to fix this issue.
However, Fiji has other problems, if we take a close look at the reviews of TechReport and Nordic Hardware: the real bandwidth of HBM is very low compared to theorical data.
From TechReport (Beyond 3D Suite): Hawaii @ 76% of Theorical Bandwidth, Fiji @ 65% of Theorical Bandwidth.
The results from Nordic Hardware are equally terrible. AIDA64's Memory Copy benchmark shows the memory (video RAM) bandwidth than a software or game can utilize in the real world. Grenada can push 78% of the theoretical bandwidth of 390400 MB/s, while the Fiji is only capable of pushing 70% of the theoretical bandwidth of 512000 MB/s.
We have 3 possibilities to explain these results:
- immature Firmware/BIOS of Fury X
- immature Catalyst driver
- HBM is a Beta product and doesn't perform as expected
HBM doesn't seem to provide considerably higher overall performance than GDDR5, and we have to overclock the HBM to pust Fiji to its limit. Also, the huge 4GB VRAM amount limitation could put AMD in a dangerous situation if we want to play at 4K with some games. Fury X seems to be an unbalanced product.
At last, but not the least, this data could explain the NVIDIA's choice to avoid the HBM and to use the HBM2 with Pascal during 2016: more bandwidth and less problems.