1. not using taa is quite short sighted, temporal upscaling gives much better visuals to performance ratio
5.
- I obviously downscale shadow res depending on settings, but simply said you cant have nice shadows for free
- with upscaling I think you are missing the point. Upscaling is to gain a tons of performance, its like upscaling from 1080p to 1440p so good sharp visuals for good loweres performance
- object cant have static shadows with dynamic lighting, and baking stuff is out of the question
- I tweak like 5 variables to get best looking game for best performance, since people are clueless and would set it wrong and complain about it later. Its a bit restricting but better this way in general
- I cant set each tree to cast or not cast shadow individually because its instance based (each tree has same settings) and it would be too labour intensive
- sending data into regular ram is the problem I'm having, it instantly reduces fps almost by half
also I'm not a rendering engineer, changing those things take few experts in industry years to do, I'm left with what guys at epic make
6. I obviously try as much as possible to avoid water rendering so every water surface futher than few meters is blended into opaque material that is quite fast to render, I just didn't mention these things here as it would be too much text
as said all my answers are overly simplified and these things go into way way deeper detail and its often not intuitive at first glance
Not offering other forms of AA is short-sighted. Not everyone has the power to run TAA and most games offer more than one form of AA. I don't know what kind of RAM usage TAA uses, but if a simpler AA consumes only 100 mb of RAM and TAA offers better visuals but consumes 400 mb of RAM, users might prefer a loss in visual acuity to the increased RAM requirements.
As you are a fairly competent programmer, you probably know that everything has a cost, usually in the form of GPU / CPU Cycles. Do too many fancy effects and you'll bottleneck the Cycles and have massive framerate loss. Also, the number of pixels rendered at 1080 vs 1440 is quite different, and as a result consumes more VRAM to do it. Especially if it is a wide screen monitor. Especially if the game was designed with 1080p in mind and now has to do approximate pixels which adds further cost. It is always more performant to build the elements and save them on the hard-drive, than to calculate them at runtime. Its like doing a pre-made animation of a cum effect vs calculating out the geometry at runtime. The pre-made stuff is always going to run more efficiently, but it doesn't have the dynamic nature of calculating at runtime, so there is always tradeoffs. Sometimes its better to have less visual fidelity and more performance, and other times its better to push performance as much as possible and try to achieve that visual fidelity.
UE4 was designed specifically in mind with baked shadows, with extremely sparingly usage of dynamic shadows and lighting effects. UE5 appears to be better for dynamic shadows, so you might have to upgrade to UE5, if your game relies too heavily on dynamic effects.
Personally I like having the options, because I'd rather have an okay-ish looking game at a solid 60 fps than a super pretty game at 20 fps. Now I might choose the wrong options and make it uglier than necessary, but, in video games I prefer framerate and gameplay stability over visuals and will always degrade the visuals if my performance is terrible. And working with a 2015 built computer, I often have to make visual sacrifices to reach 60 fps. It is a still a decently powerful rig, but it can't do modern stuff like raytracing and has trouble with games that require HDR over Bloom.
Data sent into RAM works best when it stays there for a while. If you are sending extremely transient data to it, you are sending the wrong data to it. VRAM works best with transient data. The issue is primarily heat-management and clock speed of the RAM itself. VRAM clock speed is much higher and works better at higher temperatures, while RAM is typically 500mhz or slower than VRAM, and is much more sensitive to heat, but as it isn't being accessed and deleted from constantly, its usually not a problem. Every time you send data to the RAM, it requires electricity and that generates heat. More times you do it, the more heat generated. At a raw physics level, regular RAM just cannot cope with the stresses typically placed on VRAM, so if you are sending the type of data that VRAM typically processes to RAM, you are definitely going to have a framerate loss. So, if you can think of ways of keeping certain transient data more fixed, then you could do what you are trying to do.
Think of it like this:
1. VRAM is short-term memory in a human. Usually lasts for 1 hour or less.
2. RAM is medium-term memory in a human. Usually lasts for 3 months (the typical life cycle of any cell in 3 months)
3. HDD/SSD is long-term memory in a human. Usually lasts until the cells die and fail to transfer the data over to a new cell.
