Introduction

Why?

If you ever scrolled through popular mods for most Bethesda titles, you might have noticed the sheer amount that either claims a performance improvement, or improves visuals while ignoring the performance implications (or sometimes lying about it for downloads).

Hardware & Software Used

  • HW and SW (including drivers) details can be viewed by hovering over the graph's title.
  • Data is collected through CapFrameX, capture data is available for download in the repository.
  • The mods used as a base are listed in each game's section. If a different list is used, it will be specified.

Methodology & Info

  • Don't give much consideration to 1% or 2% differences - there is always margin of error, especially when CPU bound.
  • Any outlier result is corrected by aggregating more runs and excluding anomalous results.
  • Testing the effect of a mod on CPU-side performance requires being completely limited by the CPU while profiling, this can be done with a low enough resolution and/or a fast enough GPU.
    In 99% of cases, you can tell you're CPU bound when the GPU core usage is not maxed out (99/100%).
    Refer to this video for more information.
  • Even if some of the performance results seem so high in FPS, remember that performance gains/losses largely scale the same across different hardware, obviously assuming the testing enviroment is comparable and nothing weird is going on.

    Always interpret performance differences in percentages! Absolute values should only be considered for the sake of cross comparison, since the scenes used for testing might be shared.

    For instance, telling people that a mod gives 30 more FPS while you're at 30 FPS equals to a 100% increase in performance. While it would mean only 15% at 200 FPS.

  • In any (hopefully) scientific test, all variables should be controlled as much as possible. It would be unwise to test singular mods or changes in a very complex environment such as a long modlist. For the same reason, AI is disabled during all tests unless the subject of testing is AI-related.
  • Most of the tests re-use the same save files for consistency, but some specific mods may require their own dedicated saves. The common saves will be provided for download in the repository.
  • If a test does not include CPU or GPU bound results, assume that it doesn't affect those scenarios enough to be noticeable.

FNV Benchmarks

Modlist

  • VNV's Custom INI
  • FNV 4GB Patcher
  • xNVSE
  • JIP LN NVSE Plugin
  • JohnnyGuitar NVSE (All Tweaks Preset)
  • NVTF - New Vegas Tick Fix
  • kNVSE Animation Plugin
  • UIO - User Interface Organizer
  • FNV Mod Limit Fix
  • Yukichigai Unofficial Patch - YUP
  • lStewieAl's Tweaks and Engine Fixes
  • lStewieAl's Engine Optimizations
  • Faster Start Menu
  • Testing commands (TAI, TCAI, FW NVWastelandClear)

Notes:

  • Some of the mods above will be tested individually, but the modlist will be used as baseline for the other tests.
  • DXVK is not considered baseline due to its higher variance compared to D3D9 and the many cases of incompatibility.

Engine Optimizations

UI element - Chevron

Loading benchmark data...

Conclusion and notes: For some reason it may lower FPS lows considerably, but in most cases it still improves performance enough to be worth using.

Mod links and versions:

New Vegas Tick Fix

UI element - Chevron

Loading benchmark data...

Conclusion and notes: Essential.

Mod links and versions:

DXVK

UI element - Chevron

Loading benchmark data...

Conclusion and notes: Increases CPU performance noticeably in most cases, but compatibility can vary a lot with different GPUs, drivers, and DXVK versions. Unless you're on the Linux AMD RADV drivers, it also causes dithering. All the information you'll need to use DXVK effectively is in the FNV Performance Guide.

Mod links and versions:

FO4 Benchmarks

Modlist (OG EXE + NG Content)

  • Bethini Pie High
  • F4SE
  • Plugin Preloader
  • Address Library
  • Backported Archive2 Support System
  • High FPS Physics Fix
  • Unofficial Fallout 4 Patch
  • Community Fixes Merged
  • Buffout 4 NG

High FPS Physics Fix vs Long Loading Times Fix

UI element - Chevron

Scenario 60 FPS Loading 350 FPS Loading No Loading Lock 350 FPS Loading + HFPF Speed Ups 350 FPS Loading + HFPF Speed Ups + Long Loading Times Fix
COC DiamondCityExterior 15.57s 9.91s 9.30s 9.21s 9.10s
Entering DC 20.90s 6.92s 17.20s 3.25s 4.25s
Exiting DC 23.16s 9.16s 29.91s 5.79s 6.83s

Date of testing: 2/9/24

Conclusion and notes: The speed ups mentioned are DisableBlackLoadingScreens and DisableAnimationOnLoadingScreens. Tested using the High FPS Physics Fix loading timer.

The results show that:

  • Increasing the loading screen FPS cap from 60 to 350 significantly reduces load times across all scenarios.
  • Removing the loading lock entirely breaks the loading time in most cases.
  • Applying HFPF's speed ups further improves load times.
  • Adding Long Loading Times Fix on top of HFPF speed ups does not always result in further improvements and, in some cases, slightly increases load times.
Overall, Long Loading Times Fix is made redundant by a proper HFPF configuration, as instructed here during The Midnight Ride guide.

Mod links and versions:

SSE Benchmarks

Modlist (1.6.1130 without the AE DLC)

  • Bethini Pie High
  • SKSE
  • Address Library
  • Engine Fixes
  • Bug Fixes SSE
  • Scrambled Bugs
  • Display Tweaks
  • Unofficial Skyrim Special Edition Patch
  • ConsoleUtilSSE NG
  • Autorun (TAI, TCAI, TGM, FW 81A)

Skyrim Priority

UI element - Chevron

Loading benchmark data...

Conclusion and notes: Windows already gives CPU scheduling priority to the foreground application, making most of these CPU priority mods pointless and an easy DP farm for people.

Mod links and versions:

Lightened Skyrim

UI element - Chevron

Loading benchmark data...

Conclusion and notes: As you may predict, removing a few objects in scenes with thousands of others is not going to do anything noticeable.

Mod links and versions:

Insignificant Object Remover

UI element - Chevron

Loading benchmark data...

Conclusion and notes: Grass is instanced and the CPU performance from its draws is barely relevant. This means that the mod realistically only affects tri-count (affecting GPU performance), and clearly not enough to even be measured.

Mod links and versions:

Grass FPS Booster

UI element - Chevron

Loading benchmark data...

Conclusion and notes: CPU performance isn't measured for the same reason as Insignificant Object Remover (instancing). The mod reduces visual fidelity and doesn't bring any noticeable performance improvements. The plugin is also riddled with out of scope changes.

Mod links and versions:

JK Mods

UI element - Chevron

Loading benchmark data...

Conclusion and notes: Up until the release of JK's Riften Outskirts, jkrojmal kept stating that the mod could perform at 60FPS with an ENB and who knows what else loaded on an i7 2600 and 980Ti. Considering the age of those components and the huge hit from the mod + just ENB alone, I find that really hard to believe, especially because the statement remained identical for years.

Personally, I find the interior decor of most of the suite to be quite poor taste and more aimed towards maximalist design, which really clashes against the rest of the game's lack of detail and just wants to look fancy for pictures. The performance hit is not justified at all.


Mod links and versions:

DBF LOD Guide

UI element - Chevron

Loading benchmark data...

VRAM usage:

Scenario Vanilla Terrain LOD DynDOLOD Terrain LOD + DynDOLOD
GPU Bound 1 3.03GB 4.38GB 4.96GB 5.02GB
GPU Bound 2 3GB 4.31GB 4.86GB 4.9GB
GPU Bound 3 3.86GB 3.86GB 4.49GB 4.56GB

Conclusion and notes: Regenerating terrain LOD is a no-brainer, it looks significantly better and the performance impact is minimal. The DBF DynDOLOD settings are fairly maxed out, and even then the performance impact is nothing outrageous. For most systems, the bigger concern is probably the almost 2GB (and this is almost a minimum with how vanilla oriented DBF is) of added VRAM usage when using terrain LOD + DynDOLOD with these settings, so make sure that you have enough VRAM available.

Mod links and versions:

DBF Grass LOD Guide

UI element - Chevron

Loading benchmark data...

Conclusion and notes: The tests were taken with Community Shaders (+LLF because CS by itself reduces performance due to bad vanilla lighting code) for consistency with the setup required by complex grass patches. Grass LOD is very heavy so I'd consider it only for very high-end systems and/or screenarchers.

Mod links and versions:

Tree Mods

UI element - Chevron

Loading benchmark data...

Conclusion and notes: EVT is a great option for people who want to keep the vanilla look at the same performance. Weirdly, the mod consistently lowers FPS lows in specific scenes, it's not enough to matter realistically but an interesting find nonetheless.

Mod links and versions:

Community Shaders Core

UI element - Chevron

Loading benchmark data...

Conclusion and notes:

  • Community Shaders without Light Limit Fix will perform noticeably worse than vanilla due to LLF's replacement of vanilla code that CS would otherwise fall back to. This means that CS with LLF has to be considered as the base setup for CS in order not to skew other results. This might be resolved in future versions of CS, but for now it is what it is.
  • For some reason, LLF drastically increases CPU performance variance in interiors, making profiling very tedious and often unreliable. For this reason, don't give much consideration to the percentage lows of LLF.
  • Both CS and LLF reduce percentage lows when GPU bound, although the average FPS is often improved.


Mod links and versions:

Screen Space Shadows

UI element - Chevron

Loading benchmark data...

Conclusion and notes: Screenspace Shadows is a GPU-bound effect, meaning that it will not affect CPU performance. It performs really well, and bumping up the samples multiplier improves quality without a significant performance hit[1].

  1. This can depend on the GPU architecture. As shown the 2nd graph, a 6700XT might perform marginally better with 3x samples than with 2x. Attempting to delve into something so complex would be too long winded and most likely inaccurate, so it suffices to say that various GPUs will have different optimal settings. Considered how light the mod is, it most likely won't matter much in practice.


Mod links and versions: