Current Generation Games Analysis Technical Discussion [2023] [XBSX|S, PS5, PC]

Status
Not open for further replies.
Late to the discussion but in terms of depreciating rasterization support what are we even talking about specifically in terms of the hardware side? And how much of that could be adequately essentially run in software via the shader cores as performance grows (essentially negating the need for depreciation)? Hasn't there been rumors (or patent readings?) regarding compute based ROPs?
In terms of hardware, I don't think hardware vendors are ever going to ditch texture samplers. Intel eventually figured out that it was a non-negotiable part of modern graphics hardware during their Larrabee project and have you seen how slow ASTC format emulation is on desktop graphics hardware ?

D3D12 and Vulkan still exposes warts such as tessellation, geometry shaders, and transform feedbacks. Mesh Shading isn't a true replacement to the old geometry pipeline either since it's not compatible with transform feedbacks. On some vendors you now have two internal hardware geometry pipeline paths as a result of this asymmetric functionality. Only Metal API attempts to avoid this mess at the cost of developer convenience and there's only one vendor currently that supports a near stateless approach to ordered geometry amplification ...

For things that we think of that are "programmable" such as pixel shaders, they too have "graphics state" which is implemented with specialized hardware as well that interacts with features such as MSAA, delta colour compression, optimal surface tiling memory layouts, variable rate shading, etc some of which I mentioned that we could live without ...

I don't imagine we'll ever reach a point into dropping ROP/blending hardware in the future. You get free sorting, reduced bandwidth consumption, and a predictable blending order with a relatively small amount of circuits. The alternatives like implementing per-pixel sorting in software or tile memory in hardware aren't very good. Per-pixel sorting will lead to lot's of memory traffic to the video memory in a time where bandwidth is becoming a premium and tile-based renderers (mostly mobile GPUs) aren't scalable in scene complexity for details that I'm not likely to get into once more ...
 
Not only is the performance impact huge, the UE5 doesnt support a sophisticated Raytracing system - so no standalone RTAO or reflections. There is only Lumen. And Lumen is so slow and has still so many errors. And these Lumen reflections in Fortnite 5.1 are just a joke. This software approach to Raytracing and this stacking of dedicated techniques without reusing ressources is absolut shocking.
Reflections in Fortnite trace into the surface cache only. For proper reflections you need to enable Hit Lighting.
 
It does but doesn't Lumen in general move to HWRT acceleration if it detects the hardware is there?

My understanding is Lumen defaults to HWRT off and you have to explicitly turn it on where it's available. This article has some good comparisons.

 
In terms of hardware, I don't think hardware vendors are ever going to ditch texture samplers. Intel eventually figured out that it was a non-negotiable part of modern graphics hardware during their Larrabee project and have you seen how slow ASTC format emulation is on desktop graphics hardware ?

Texture samplers aren't raster specific though. They're universally used for filtering data and coloring pixels in raster, RT and compute.

don't imagine we'll ever reach a point into dropping ROP/blending hardware in the future. You get free sorting

Does ROP hardware implicitly enforce sort order or is it the api?
 
That can’t really happen since GPUs don’t spend a lot of time actually rasterizing triangles. Most of the time is spent on shading/compute and moving buffers around. Any future GPU would be faster at compute, memory access etc and therefore still win.



The more accurate analogy is that today’s games give you some real good blow up doll action and you love it because you’ve never been with a real woman.
We live in a world where new gpus exist that perform worse than older gpus…. Not only can it happen, it happens frequently due to vendor product segmentation. The point I was essentially trying to make is that GPU vendors cannot completely abandon the pursuit of increasing raster performance to chase raytracing.


Also, your analogy is kinda off. A blow up doll doesn’t nag, doesn’t argue, doesn’t throw tantrums, drain my bank account, isn’t moody, doesn’t look worse as they it ages, etc… Path tracing is supposed to be superior right? Anyway, I digress.
 
Last edited:
The alternatives like implementing per-pixel sorting in software or tile memory in hardware aren't very good. Per-pixel sorting will lead to lot's of memory traffic to the video memory in a time where bandwidth is becoming a premium and tile-based renderers (mostly mobile GPUs) aren't scalable in scene complexity for details that I'm not likely to get into once more ...
I'm only familiar from the user-land code side, not the hardware or the driver, but mobile tile based culling/sorting seems quite good -- do you have a resource on those details?

Also, your analogy is kinda off. A blow up doll doesn’t nag, doesn’t argue, doesn’t throw tantrums, drain my bank account, isn’t moody, doesn’t look worse as they it ages, etc… Path tracing is supposed to be superior right? Anyway, I digress.
yeah dude we get it you hate raytracing and women, can you grow up/move on.
 
Very brief RE4 Demo report, both PC (RTX 3060, 12400f) and PS5:

First off, PC: Woot, doesn't look like they fucked it up! No DLSS (boo), but FSR2, so goddamn finally something better than interlaced mode.

It's demanding alright compared to Village, on my GPU without RT, you need FSR Quality at even 1440p to maintain 60 (most settings high/max, just SSAO though). 4k with FSR performance will get you ~50fps outdoors but can just barely manage 60 indoors. So without RT, I'd say you'd have to start at a 3060ti to run 4k/60 with FSR performance. FSR seems decent at least, hair is a little frizzy but the image is stable.

So no disaster like Village at launch at least. REngine games have never really had a problem with shader stutter mind you, but of course Village had other stuttering issues that this doesn't exhibit. I even mistakenly installed it to my 5400rpm HDD and no traversal stutter to speak of other than just longer initial load, so that's surprising.

Some longstanding annoyances are finally rectified too - such as changing the resolution no longer actually changes the resolution with each slider change, only changes it when you select to confirm it. Selectable PS4/Xbox glyphs too, nice.

PS5 Version: Has a Prioritize Frame Rate and Prioritize Resolution mode. Both aren't particularly sharp, and prioritize frame rate seems to exhibit more specular shimmering than on my PC at 1440p with FSR Quality, so maybe a lower mode of FSR or they're using their own checkerboarding, not sure. It's a locked 60 at least. Prioritize Resolution is clearly sharper but doesn't look like 4K, however it's an unlocked framerate, which means at least in the starting area it's in the high 40's-low 50's (?), so unless they add a 30fps cap option probably best to avoid it without a VRR display.

Both modes allow you to enable RT, but again though it's just a sub-60fps unlocked experience even if you choose prioritize framerate. So the same dumb decision they did with Village and RE2/3, don't get why they just don't lower the res further or just have a 30/40fps cap option too - or maybe that might show up on a 120hz display. Either way at least on my 60hz display, only prioritze frame rate with no RT seems to be the viable option atm. All modes require you to exit the game and go back to the main menu to change, annoying when most games can change res/frame rate modes on the fly.
 
My RE4 demo report is that it is phenomenal on PC. Game is so stupidly butter smooth, and looks so good. Tons of options to tweak as well with a few new ones added to the mix.

The ONLY issue I have with it... is that cutscenes have black bars in Ultrawide.

They'll probably be removable easily enough... but yea.
 
I don't really get why they don't lock to 30fps or 40fps on the more demanding modes. Well I guess it works for future proofing but still 😂
Yeah a balanced 40fps mode should have been an option for capable displays by default
 
Texture samplers aren't raster specific though. They're universally used for filtering data and coloring pixels in raster, RT and compute.
Texture samplers use pixel shader derivatives to perform mipmap selection. The graphics pipeline is intricately linked with each other along the stages ...

Even the resource binding model shows performance side effects on pixel shaders as an example ...
Does ROP hardware implicitly enforce sort order or is it the api?
Virtually all 3D gfx APIs mandates rendering to be in primitive order ever since it's inception. ROPs are one of the few ways to enforce primitive sort order. Hardware and software are built and designed to match the requirements for each other ...
I'm only familiar from the user-land code side, not the hardware or the driver, but mobile tile based culling/sorting seems quite good -- do you have a resource on those details?
Each tile-based renderers have their own somewhat unique implementations ...

On Apple/IMG, there's a parameter buffer which has a limited amount of memory space. If it fills/overflows the driver can then either flush it which will consume more video memory bandwidth and will not be able to get the full benefit of the hidden surface removal technique or the driver can choose to do a crash/device loss which isn't all that helpful ...

On ARM, they have a simultaneously split geometry pipeline which is downright incompatible with geometry shaders since it makes their tiling/binning phase redundant ...

In both cases, sorting occurs before fragment shading. This sorting combined with tile memory makes it trivial to implement programmable blending hence the description of a tile-based renderer or less commonly known as "sort-middle" architecture since sorting happens right in the middle of the graphics pipeline post-geometry/pre-fragment shading ...

With immediate mode renderers, sorting occurs during the last stage of the graphics pipeline when we go about blending/compositing our image for output and is therefore referred to as "sort-last" architecture ...

This sorting classification has fundamental consequences. Pre-pass techniques become useless on tile-based renderers since sorting earlier in the pipeline allows hardware to never see overdraw during fragment shading while the opposite is true for immediate mode renderers. Sorting later in the pipeline has it's own benefits such as having a less restrictive geometry pipeline (first implementations in T&L/vertex/geometry/tessellation/transform feedbacks and now mesh shading) and more parallelism being available during fragment shading. Deferred rendering on immediate mode architectures is one of the more powerful combinations that we commonly see out there in high-end graphics since we can separate the material/lighting into separate passes which lowers register pressure thus letting hardware launch more invocations leading to increased occupancy. Occupancy or register pressure isn't as big of an issue on tile-based architectures so a forward renderer with no prepass becomes the optimal composition. On immediate mode architectures, the cost of another geometry pass is starting to become trivial so the ideal fusion is a deferred renderer with a visibility buffer pre-pass since the geometry pipeline is more scalable to handle higher scene complexity and can exploit more parallelism by design ...

One could make a robust argument that modern graphics technology has evolved the way it did because of these origins ...
 
Forspoken and Hogwarts received a patch that vastly improves texture quality, performance and VRAM usage. 8 GB GPUs redeemed.
I wouldn't say redeemed but more like on life support. Sure it is fixed for now but for how long the treatment will continue to what extent?

After all, 2 GB GTX 770 ran Rise of Tomb Raider and Arkham Knight with great visuals with its tiny 2 GB buffer. With games after 2018; have to reduce textures to a point games look really worse than these two games I've listed due to abhorrent textures. (RDR2 for example looks visually worse than Rise of the Tomb Raider with low textures. Yes, it has better GI, shadows etc. but clearly a better "texturing" solution could've been procured. But at that point, who cares for 2 GB GPUs anyways? Sooner or later, if enough people move on from 8 gig cards, devs will stop doing these "patches". It just proves that capability and ability and doability is there. It just depends on the user outrage. (Though in the case of Forspoken, ship has sailed already. However if it indeed fixed the problem, I could consider getting the game on %50 sale)
 
Status
Not open for further replies.
Back
Top