Do GPUs really need 4gb of RAM?

[bm23]

It seems that everyone is clamoring for at least 4gb of VRAM on GPUs, but from my experience, VRAM has not been so crucial (my stats below). I understand that many people game at higher resolutions than 1080p, but even at 5760x1080, GTX 770 2gb SLI is still doing a damn fine job (LINK). So, my question is, are gamers too caught up with VRAM? How much difference would it make in real world gameplay? The reason I'm asking is that this rush for more VRAM doesn't quite add up with benchmarks. Perhaps I'm missing something. Either way, maybe this discussion can prove useful for those who feel pressured to upgrade to GPUs with more VRAM.

My current rig:
GTX 770 2gb SLI
i7 5960x
32gb 2400mhz DDR4

Games fps:
Shadow of Mordor: 30 @ 4K, ultra | 50-60 @ 3K, ultra | 85-95 @ 1080p, ultra
Witcher 3: 35-45 @ 2715x1527, ultra, NVidia Hairwork Low | 45-60 @ 1080p, ultra, NVidia Hairwork Low
Tomb Raider: 75-100 @ 2715x1527, ultra with TressFX
GTA 5: 60 @ 1080p (mostly maxed out, except for those few advance settings. I haven't played in a while so I don't remember exactly)
Metro Last Light: 40-45 @ 1080p, maxed out
Dragon Age Inquisition: 40-50 @ 2715x1527, maxed out | 70-90 @ 1080p, maxed out
Dying Light: 70-100 @ 2715x1527, maxed out

 

[beers]

There are a bunch of reviews around comparing various cards at 4 GB vs 2 GB. Depending on the game there's less of a minimum FPS nosedive on cards with higher amounts of VRAM.

 

[Darren]

GTA V I know will use more than 2GB pretty easily if you have the VRAM. My usage is about 2.3GB out of 3GB and I'm not even maxing it. The R9 390 has 8GB of VRAM already, although I'm not sure when that will be necessary.

 

[Shlouski]

If you are running a game at higher settings that needs more VRAM than is available, then the card will start to use system RAM and slower RAM will decrease performance significantly. In GTA5 I hit a maximum of 3.5gb of used vram, but there is no point in having lots of vram unless you have a good enough gpu to process all that data.