I said this before, but I am not convinced that storage bandwidth should be weighted equally as CPU/GPU raw render ability in the overall rendering process
This right here. They are apples to oranges overall.
Also, you triggered me into going into a wall of text, but I got some examples for this stuff. Those with Wall-of-Text-phobia have been warned
......................
(
NOTE: I use LOQ as "Level of Quality"; I don't necessarily mean it the same was as LOD, or Level of Detail, although there's some overlap. LOD is, per Wikipedia: "involves decreasing the complexity of a
3D model representation as it moves away from the viewer or according to other metrics such as object importance, viewpoint-relative speed or position. Level of detail techniques increase the efficiency of rendering by decreasing the workload on
graphics pipeline stages, usually
vertex transformations."
I'm using Level of Quality/LOQ as a means of marrying that with visualized parameter boundaries specified in virtualized game-world metrics, such as feet (FT), and arbitrarily stating different LOQs for various ranges of feet (FT). Just so things are clear on the terminology that follows.)
Maybe if people pictured it like a 2D camera system it would make a bit more sense. Like, in a lot of 2D games, you actually have a camera system (of sorts), because there are still objects and visuals being calculated outside of the immediate viewing frame, just possibly at a lower rate.
The frame window is the GDDR6 RAM, and let's say the data outside of it is what comes in from the SSD. Let's split that up into a few rings. The immediate ring around the main frame window's perimeter is the Level 1 priority data, it's the data to immediately replace and/or swap assets in RAM (if it needs to; otherwise I don't see why the SSD can't simply stream this through to the GPU if there's a means to do so which I
believe is the case for both PS5 and XSX). So say the frame window is a "L1 cache", then this first "ring" would be the L2 (I am
NOT saying these things are actually analogous with system memory caches, just saying that if you wanted to give a point of comparison to better visualize things, you could structure a visualization similar to system memory caches)
Let's say there are several other "rings" (consider these rectangular rings) around that first ring; all of these represent data on the SSD being streamed in at different levels of priority. And let's say the "rings" have different levels of quality (I'll just call it LOQ). The immediate LOQ (the one with the active frame window, representing the RAM) is LOQ 0/zero, the highest quality. And let's just arbitrarily say the "rings" surrounding that have a linear ranking from 1 through 8, 8 being the lowest quality level.
The proportion of each of the 1-8 LOQs would change depending on a lot of different factors; not every game will have the same demands, and this is influenced by things such as design style (linear level-based, open-world, etc.), graphical styles/techniques, game design choices, implementation of game mechanics and physics systems, etc. By and large, let's just take the general raw speeds on both PS5 and XSX to define a ratio of 2.25:1. In other words, for all the rings 1 through 8, PS5's would be 125% larger, but if we're talking, say, wide open spaces of thousand of feet in span, and you get an average of 1000 FT per "ring", on XSX LOQ 1 would be 1000 FT while on PS5 it'd be 2250 FT, just as an example. And this is assuming that the sizes/proportions per "ring" would stay the same/be even across the board (i.e the sizes are a fixed constant of the same value for each LOQ).
The question would then become, at what "ring" level does depreciating LOQ ratings
not have any perceivable impact on a player's experience? For example, if in the aforementioned example PS5's LOQ 1 "ring" is 2500 FT and XSX's LOQ 1 "ring" is 1000 FT, how much does that
actually impact the player's experience in terms of immersion? Because let's also keep in mind, the LOQ 0/zero "ring", which is the one the RAM space occupies, that's the immediate visual data directly around the player (measured in terms of virtualized game real estate), and if we assume a 1000 FT LOQ ring is equivalent to, say, 10 GB of physical data on the SSD drive, then we can assume LOQ 0/zero can provide, let's say, 250 FT of the highest-level asset quality (I reduced it from 1000 FT because what's in RAM I would assume is also being ran through intense physics and interactive simulations by the player, and also various other scripting logic, character model assets, enemy models, etc.).
So now let's ask ourselves, is there really a
perceivable (as in, during
actual gameplay, seen by the average player) gulf in quality if one system can produce highest-quality and extremely high-quality asset texture quality in a virtualized game space of, say, 1250 FT at any one time (XSX), vs. 2500 FT at any one time (PS5), if specifying a cutoff at LOQ ring 1, and assuming this is an actual game with true logic and physics, AI, and enemies happening on-screen? Will the
average gamer be able to tell some tree at a distance of 1250 FT from them has any glaring difference in visual texture quality than a tree literally a
foot in front of them?
Chance are, probably not. Firstly because what average gamer is going to be paying attention to that tree
far off in the distance? More importantly, if that happens to be the case, you then have to ask is it even
worth having a high-quality asset texture model upwards even 2500 feet away from the player if they will not be viewing said model at the given moment? Maybe you can get away with a lower-quality version of that model at that distance after all and there is no perceivable impact to the visual immersive experience for the average player, considering most real-world gameplay scenarios. It
does require a bit more work from the visual artists I suppose, but I figure advancing toolsets would automate much of that type of stuff for them regardless.
If that happens to be the case, a developer can basically go "well, we really don't
need high-quality texture assets for these objects at
distance, so let's make LOQs at 4 and lower (as in, 4, 5, 6, 7, and 8) 100 FT instead of 1000 FT." Okay, cool. Now that frees up a bigger range for high-quality texture assets in LOQs 1, 2, and 3. Now if that means LOQ 1 goes from covering 1000 FT to, say, 2250 FT, for the XSX that would effectively put its LOQ 1 at PS5's default LOQ 1, though it means the XSX had to drop the range of lower LOQ rings notably to do so.
The thing is, in
many games (in fact I'd say the vast majority), this would be perfectly fine, because again just how far off into the distance are we
really expecting MOST gamers to painstakingly pay attention to when it comes to details? More importantly, if you're familiar with classical painting techniques, you'll know that you can actually hint/indicate an
impression of detail without actually providing much or even
any detail in a given area. This is usually done by utilizing smart color pairings and value contrasts; in fact that's exactly what we saw in the UE5 demo in areas such as on the girl character's skin with the light coming down as she crawled through a cave passage.
Now, everything I just mentioned perfectly applies to PS5 as well and say, for example, a dev could lower the range on the lower LOQ "rings" to bump up the range on the first upper ones. So let's say it's by the same rate as the XSX example; that could then give the PS5's LOQ 1 "ring" a range of 3500 FT instead of 2250 FT. Again, for
certain types of games and
some real-world gameplay scenarios this can be beneficial, but there's
always a point of diminishing returns because by and large players will focus on the details immediately in front of them and immediately around their avatar.
So I can actually also switch over to talking about this a bit in terms of simply increasing the level of detail on immediate (LOQ 0, LOQ 1) texture assets, as well. Because like I just said, not every game needs intense texture detail on objects super far-away (and in fact, sometimes you can get very comparable or superior results with smart utilization of various visual effects and low-quality model assets to give impressions of details that work just as well, and save a ton on the data streaming pipeline). Again, picture painting techniques from the Old Masters
Now, the issue with this is, at
some point you're
going to hit a wall in terms of the limit you can push the texture asset quality being streamed in vs. what the system can
ACTUALLY output in terms of resolution to the display device, because while the
former is reliant on the
SSD, the
latter is reliant on the
GPU (and if you want to go a step further, the fidelity of those two would be reliant on the CPU).
What's the point of trying to pump out a 16K texture asset to stream from storage when the game resolution is going to be "only" 4K (and that's just in some cases; we're probably looking closer at 1440p - 2160p for a good deal of next-gen games, tho I'd like 4K60 to be more of a standard personally)? The
perceived level of difference to the average player between an 8K version of that texture and a 16K version of the same texture at an output resolution of barely 4K on a 4K screen, will likely be imperceptible. After all, if most people already have a hard time telling much of any immersion-breaking (or even immersion-irritating) difference between things like RE3 Remake on PS4 Pro vs. RE3 Remake on the One X, that argument
should perfectly carry over to native texture resolution asset streaming that reaches a point where the returns rapidly diminish on most standard modern 4K displays, right?
Again this is where I'd like to focus on things like DLSS techniques for a moment, because at least IMO, if you had the choice between streaming through extremely large, super-high resolution 16K or even some 8K texture assets from storage into memory, or from storage through the GPU via some implementation of an AMD SSG-type setup (which I believe both systems are doing TBH), versus rolling with a lower-quality texture you can simply upscale to a higher resolution that appears pretty much like it were a native resolution anyway...I would think most would choose the
latter. After all, it saves on game disc and game install space, file size, and that frees up the SSD data pipeline for moving through more unique data overall. In particular for artists who like to use some notable degree of programming to create their art, the latter approach would probably be their preferred one (granted it's not the only approach, of course).
I thought that was gonna be a shorter explanation than it turned out to be
, goes to show how complicated this stuff is. It may not be a
perfect explanation (is anything ever, truly?), but I had an interest in trying to visualize a breakdown of how PS5 and XSX's SSDs could operate on a basic level with some practical examples and considering majority of real-time, general gameplay use-cases and gamer habits. And also, to show how you don't always need "more" to accomplish "more", depending on a variety of other factors. Maybe I'd like to elaborate on this in the future sometime, but I'm gonna wait until we get more system info from MS and Sony before doing that, so it's definitely gonna be well into August at earliest.
