Pretty much completely CPU/platform limited, even at 4k (via DLDSR +33% smoothing on a 1440p display, to help with aliasing) Ultra+, with well beyond ultra texture (8k planet and skybox, 1k environment maps) and shadow quality (16k, nine frustums, as described here).
I expect this would likely still be appreciably limited by things other than the GPU even with a well-tuned 13900KS or 7800X3D setup, though I imagine they'd still perform a fair bit better. That said, I'm finding my current performance acceptable and these settlements, plus the associated AI load of a high intensity CZ, are about the worst case scenario for CPU/memory bottlenecks. It is amusing to note that this system consumes about 200 watts less in a CZ than sitting around in a starport concourse because of the discrepancy in GPU load.
On a related tangent, I'm pretty impressed with the new AV1 encoder on these cards. This recording averaged almost 95Mbps, which seems high, but it's still quite a bit smaller than H.265 would have been at the same QP setting. Will need to find someone with a 7900 XT(X) to compare the quality vs. bitrate against with OBS 29.
If you DM me the information on what you'd like to see in the comparison, I'd be willing to give it a go. (I can even transplant the 5800 X3D into my PC to compare apples to apples)
Hard to get around loading pauses, which is the only time the dips last long enough for the game's FPS counter to register it. However the game actually dips to ~30 fps quite frequently...that's what those P0.2% framerates are. During a high CZ at a large burning settlement, fully one frame in five hundred is ~16ms long.
I normally pay most attention to the P1% frame rates as that's most reflective of the smoothness (or lack there of) I'm actually perceiving the overwhelming majority of the time.
I am also running 'ultra' low latency mode in the NVIDIA driver and this could be a contributor to weaker minimum frame rates. A deeper queue smooths stuff out at the cost of latency. A shallower queue minimizes latency at the expense of some degree of frame rate smoothness. I'll have to do some more testing.
If you DM me the information on what you'd like to see in the comparison, I'd be willing to give it a go. (I can even transplant the 5800 X3D into my PC to compare apples to apples)
The GPU encoder performance should not be at all CPU dependent, within reason. Trick would be recording identical scenes with identical encoder settings as possible then comparing the size of the files produced. At the same CQP the videos should look more or less identical, so the smaller file wins. Traditionally NVENC has had a significant edge over VCE/VCN/AMF here, but that gap has been closing fast this last year or so and both AMD and NVIDIA's AV1 encoders are new to this generation.
Elite: Dangerous isn't well suited for this as the only 100% predicable (automated) path I know of is at the end of the suit tutorial. Would probably need to record a canned benchmark for good accuracy. I'll think of something.
In the meantime, can you show me what this page looks like for you in the newest version of OBS when you have the AMD AMF AV1 encoder selected?
Pretty much completely CPU/platform limited, even at 4k (via DLDSR +33% smoothing on a 1440p display, to help with aliasing) Ultra+, with well beyond ultra texture (8k planet and skybox, 1k environment maps) and shadow quality (16k, nine frustums, as described here).
I expect this would likely still be appreciably limited by things other than the GPU even with a well-tuned 13900KS or 7800X3D setup, though I imagine they'd still perform a fair bit better. That said, I'm finding my current performance acceptable and these settlements, plus the associated AI load of a high intensity CZ, are about the worst case scenario for CPU/memory bottlenecks. It is amusing to note that this system consumes about 200 watts less in a CZ than sitting around in a starport concourse because of the discrepancy in GPU load.
On a related tangent, I'm pretty impressed with the new AV1 encoder on these cards. This recording averaged almost 95Mbps, which seems high, but it's still quite a bit smaller than H.265 would have been at the same QP setting. Will need to find someone with a 7900 XT(X) to compare the quality vs. bitrate against with OBS 29.
They look more similar than I was expecting. Anyway, that's just what the OBS plugin exposes, both have a lot more options, and you can actually enter them into the AMD AV1 plugin without having to switch to the custom FFmpeg settings.
The small (CPU/memory related) performance issues I experienced weren't obtrusive and the GPU (brute forcing resolution, plus DLDSR) helps aliasing just enough to make that acceptable. Fairly ridiculous what hardware it takes to get to this point...as I was here in Horizons with an i7-5820K and a GTX 1080 Ti, but I'm finding the end result enjoyable.
I'm also astonished at how much detail YouTube decided to retain. I'm having difficulty distinguishing my source video from the 4k60 YouTube stream at first glance. It also only took them a few hours to transcode it. Not that long ago they would have mangled a video that matched their recommended upload settings more closely and taken three days to do it.
Use as reference only, unless you have a VRR display attached to an Ampere or newer NVIDIA GPU on a platform that allows ReBAR. It enables DLDSR, sets ultra low latency mode, fast sync, forces 16x AF with maximum texture filtering quality options, and enables a custom 4GiB ReBAR size (default is disabled for EDO and 1GiB for most NVIDIA whitelisted titles). I've also disabled Force CUDA P2 state to keep the VRAM on NVIDIA GPUs from downclocking while running NVENC, which singificantly reduces it's percieved overhead.
I doubt the LODDistanceScale of greater than 1.0 is working and directional shadows are set to LOW because that's the preset I've replaced with my custom shadow tables.
Only really relevant because I've set ship and on foot FOV to exactly 60 degrees for both, which is a small bump for on-foot and the maximum default slider for the ship. It's possible to go much higher, but I find 60 entirely sufficient and more fish-eyes the crap out of the view, which bothers me.
My override, which contains the bulk of my custom settings, other than shadows. Do not use on low-end GPUs without adjusting GalaxyBackground, GalaxyMap settings, at the least.
Mostly experimental adjustments to Thread and Network that need to be evaluated further. Some of the former are probably placebo while some of the latter are depreciated. Only settings that I might recommend from here are the increases to render thread stack and queue size, plus perhaps, number of worker threads, depending on how many cores one has available (but never fewer than two or more than twelve).
XML:
<?xml version="1.0" encoding="UTF-8" ?>
<AudioConfig>
<Latency>
<RefillBuffers>3</RefillBuffers>
<SamplesPerFrame>1024</SamplesPerFrame>
</Latency>
<AudioAPI>1</AudioAPI>
</AudioConfig> -->
<!--
This file gives you the option to overwrite Elite's internal settings for the audio engine Wwise on Windows systems.
Please note that the game will sanity check all supplied values and fall back to defaults in the case of invalid data.
1. Latency:
The numbers RefillBuffers and SamplesPerFrame under "Latency" define the internal latency and thus the frame rate of the audio engine.
A rule of thumb here is that lower values mean LESS latency and therefore a FASTER frame rate.
Lowering the numbers here might be desirable e.g. when you want to get audio frame rate closer to the target frame rate of VR headsets (90fps, ~11ms).
However be aware that this will also increase the CPU impact and you might experience audio dropouts if the numbers are too low.
RefillBuffers: default is 4, valid numbers are 1 to 8.
SamplesPerFrame: default is 1024, valid entries are 256, 512, 1024 and 2048
For more info check: https://www.audiokinetic.com/library/2015.1.5_5533/?source=SDK&id=soundengine__events.html (Latency Between Posting an Event and Actual Sound Playback)
Feel free to experiment with those numbers and maybe let us know via the forums how they work out for you.
2. AudioAPI:
This setting allows you to overwrite the used Audio API on Windows systems. These are defaults:
- If you start the game within a VR environment (Oculus Store or the Steam VR build) it will use WASAPI.
- If you start the game within Steam (non VR) or via Elite launcher either WASAPI or XAudio2 may be chosen (it is Wwise's choice and depends on the used operating system).
Changing this might be useful if you experience Elite taking exclusive ownership about Windows' audio environment or silent Oculus Rift headphones.
The values mean:
0 - no change, the default AudioAPI will be used (see above)
1 - force WASAPI
2 - force XAudio2
In order to try out, simply uncomment the XML code at the top of the file.
-->
The only changes I made here were to force WASAPI and reduce the number of RefillBuffers to three, which slightly improves audio latency.
The segment from main GraphicsConfiguration.xml which has my custom shadow (detailed in my other thread, which has the full file attached) and some cleaned up formatting/tab stops. These tables cannot go in the override because they add new attributes, rather than exclusively replacing existing ones.
I'm also setting the EliteDangerous64.exe executable to use large memory pages, which seems to reduce overhead slightly. This requires running the game as administrator and modifying the Image File Execution Options in the registry. It's another one of those experimental tweaks that I need to test more throughly to verify if it's actually helping anything.
Neo G8 is the next big upgrade (assuming I can't find a better 32"+ 4k240+ display when the time comes), though EDO jaggies like DSR smoothing, so it would probably be a downgrade for this game.
