4GB VRAM isn't cutting it anymore says AMD

#5799130 Posted on: 06/13/2020 09:47 AM

---

There already is lots of compression with...
*DDS Direct Draw Surface graphics - which depending on if you need a lot of varying levels of translucency (takes double the space of) or simple single bit transparency flags (which halves the size of the file, at the cost sometimes of black outlines a'la Sonic The Hedgehog if you decide to later need it in translucent applications (it's okay for a brick wall though, that technically won't have translucent portions). Also, graphics (such as various color asphalt, or painted brick walls) can share all the color/diffuse, normal, specular/reflective or alpha maps (saving 60mb~80mb or so for a 4k texture, for every texture spared) and get colored by shaders on render (at negligible processing cost).
*Layered textures (such as creating derelict versions of asphalt, concrete, and brick for run-down portions of town, layering grunge/grime only as-needed) also can get the most out of your graphic files.
*Compressed DAE (Collada models), which can make plain-text Collada models and compress them sometimes into 1~10kb from 100~1000kb or more. Many complicated models can be larger, some much larger. You can go through 100mb of VRAM just allocating space for one single high-detail block of city without batting an eyelash if it's a modern game, lower levels-of-detail use exponentially less.
*Additionally, models can be colored in an instance-based fashion, so say you have a bunch of cookie-cutter houses (such as townhouses or newer housing developments) where they all look the same other than color difference and lot furniture... you can just color these at runtime/render, saving these from being otherwise pointless duplicates wasting VRAM space. This is separate (or at-least can be) from just coloring the textures themselves in the texture atlases in the game files.
*Baked light maps can take up A LOT of space, now we're getting to the ability to render dynamic lights (in the last 13~14 years or so) over things vs just using a generic alpha-enhancing mask (like the flashlight in half-life original, IIRC), so we don't have to leave the lighting residing in the drive, we can do this at render time. Sure it takes more processing power, but you always have a balancing act between more processing power required of the GPU and CPU and how much VRAM is getting munched up. Ray Tracing is now adding a new higher-tech more accurate way to do light maps in real-time - this is good as the old faked ways have lots of flaws / hacky methods that must be 'worked around' by the level/scene designers / programmers.
So yes, there's ways to lower use in software, before it even gets loaded into the hardware. The hardware in nVidia and AMD both have varying sorts of memory compression, which helps speed things up sometimes considerably.
I know this was much more drawn out than needed be, but it's here if anyone's interested, and definitely appropriate to the subject.
As a content creator, I can say 4gb isn't - nor 8gb - enough if you want a truly immersive environment... 16gb~32gb would be a good starting point unless you like everything repeating in the background like an old Hanna-Barberra cartoon.
We'll get there one day soon.