THANKS FOR THE MEMORY
WHY NVIDIA’S STINGY MEMORY BUSES AREN’T JUST BAD FOR BANDWIDTH
Just enough. That’s been Nvidia’s approach to graphics memory for a while now. Previous-gen Ampere cards are a good example. The RTX 3080 got 10GB of VRAM. That was just enough at the end of 2020, but looks increasingly marginal today.
Ditto the 8GB bequeathed to the RTX 3070 and RTX 3070 Ti models. It was just enough two years ago—not so much today. The problem is the number of games that use more than 8GB of graphics memory even at 1080p, let alone higher resolutions, is definitely on the up. Notable holders of the dubious honor of breaching 8GB at 1080p include TheLastofUs, HogwartsLegacy, Resident Evil4, APlagueTale:Requiem, TheCallistoProtocol, and several more.
Some would argue that those games are just poorly optimized, and that there’s no need for any of them to breach 8GB of VRAM at 1080p. But that’s by the by. The fact is they do, and the consequence for any graphics card with just 8GB of VRAM is catastrophic in terms of performance and sometimes image quality, the latter because some games simply don’t bother to load certain textures when the GPU runs out of memory. Ugly.
Ironically, it’s often enabling ray tracing that pushes the latest games well over 8GB. The result can be the counter-intuitive situation where enabling ray tracing makes a given Nvidia card slower than the AMD competition, which certainly is not normally the case. Whatever the case, Nvidia’s problem going forward is that it has been aggressively lowering the specs for its new RTX 40 series boards. That includes using smaller, cheaper chips with narrower memory buses, tier for tier, so the new RTX 4070 and RTX 4070 Ti get 192-bit buses, down from 256-bit on the RTX 3070 and 3070 Ti. The upcoming RTX 4060 and RTX 4060 Ti are expected to run 128-bit buses.
The usual complaint is that there is reduced memory bandwidth, and that will be the case, albeit offset by increased on-die cache memory. But the real problem with those narrower buses are the limitations it puts on VRAM allocation. Graphics memory chips are only available in certain densities and bit widths. Long story short, for a given GPU memory bus width, you only have certain options regarding overall VRAM amounts.
So for a 256-bit bus card, the realistic options are 8GB and 16GB, and for a 192-bit card it’s 6GB or 12GB. For a 128-bit board, Nvidia is effectively going to be limited to 8GB for the RTX 4060 and 4060 Ti. It couldn’t do 12GB even if it wanted to, and it won’t matter how much on-die cache the RTX 4060 and 4060 Ti have if they run out of VRAM playing the latest games at 1080p. Surely 1080p is not too much to ask of a graphics card that in the case of RTX 4060 Ti is probably going to cost upwards of $400. Has Nvidia gone too far this time in trying to cut costs? Watch this space. There’s a real risk of some very ugly benchmark results when it launches its next RTX 40 series cards.