Benchmarks

Benchmarks

In order to showcase the progress between different DXVK-GPLALL versions, improvements that were made to upstream DXVK's default FPS Limiter and the reason why DXVK-GPLALL 2.6.8-1 default mode was changed from dxvk.framePace = "low-latency" to dxvk.framePace = "max-frame-latency", it has been decided to provide some benchmarks. Also, benchmarks show the direction of improvements in both upstream DXVK and DXVK-GPLALL.

System information

• OS: Windows 10 21H1
• CPU: AMD Ryzen 9 5950X (Fixed Clock - 3.85 GHz, SMT Off, 16C/16T)
  • Note: Fixed Clock makes CPU Load metric more useful, because in Fixed Clock mode CPU Load metric shows the exact amount of used CPU Time at specific CPU Clock, which allows to showcase the efficiency improvements in software. Also, it negates the effect of ambient parameters on CPU performance - CPU Clock does not fluctuate.
• GPU: AMD Radeon RX 9060 XT 16 GB (Power Limit: -30%, Voltage Offset: -100 mV, Fans Mode - Fixed RPM @ 54% = approx. 1850 RPM)
• Driver:
  • AMD Software: Adrenalin Edition 25.9.1 WHQL;
  • Vulkan Headers 1.4.315;
  • Vulkan Driver 2.0.353;
  • VulkanCapsViewer_4.02 Report - AMD Radeon RX 9060 XT.json
• Display: 1920x1080, 60 Hz, No VRR/FreeSync
• DXVK-GPLALL versions:
  • DXVK-GPLALL-GCC-WinMacLinux-SSE2-O3-LTO.2.7.1-4 - DXVK 2.7.1 with some of the GPLALL improvements;
  • DXVK-GPLALL-GCC-WinMacLinux-SSE2-O3-LTO.2.6.8-1 - DXVK-GPLALL 2.6.x branch's latest release, with all available GPLALL improvements at the moment of the release.
  • Note: GCC builds are used, because GCC builds are the default builds that are provided by upstream DXVK and DXVK Low Latency.

Test Conditions

• Reference Point: Idle Desktop;
• Test Application: Borderlands 2 - Remastered (1.8.5 - SteamDB 3902911, D3D9 x32, Unreal Engine 3) - D3D9 workload, which is perfect to showcase DXVK-GPLALL 2.6.8-1 improvements and improved FPS limiter;
• Test Scene: Main Menu.
  • In Borderlands 2 in main menu in-game camera moves slowly around player's character in fixed pattern and shows distant landscapes, which guarantees consistency, repeatability, and similarity to in-game workload.
• Test Duration: 2 minutes.
• Measurement Software: GPU-Z v 2.68.0, Windows Task Manager, DXVK HUD.
• DXVK-GPLALL and Tested Application Settings:
  • All hardware/software and system/in-game settings are the same for respective apllications between tests.
    • Note: In-Game Graphics Settings are not provided, because they are modified through application's config files to get better performance/visuals and, therefore, do not reflect actual graphics settings.
  • FPS Limit: 30 FPS (due to 60 Hz Display), set in dxvk.conf.
  • V-Sync: On, set in applicaiton.
  • Every DXVK-GPLALL version was checked in two modes:
    • dxvk.framePace = "low-latency" - DXVK Low Latency and DXVK-GPLALL up to 2.6.7-1 default setting; uses DXVK Low Latency FPS Limiter and tuned to get lowest render and input latency at the cost of the performance.
    • dxvk.framePace = "max-frame-latency" - upstream DXVK default setting; uses upstream DXVK FPS Limiter and tuned to get maximum performance - default setting in DXVK-GPLALL 2.6.8-1 and later
  • Every other setting for every respective DXVK-GPLALL version was in its respective default state for each test.
  • Every run is a 1st run: state cache and shader cache deleted between runs.
• Ambient Conditions: Same between tests.
  • Ambient Temperature: +27 degrees Celsius.

Performance Metrics Comparison - CPU Load (%), CPU Temp. (C), GPU Load (%), GPU Temp. (C), GPU Watts, RAM Usage (MB), VRAM Usage (MB)

• Idle Desktop - Reference Point:

Idle Desktop	Value
CPU Load (%) - Average - Task Manager	0%
CPU Load (%) on Cores that run Application - Task Manager	-
CPU Temperature (C) - Average - GPU-Z	39.0 degrees Celsius
CPU Temperature (C) - Highest - GPU-Z	39.1 degrees Celsius
GPU Load (%) - Average - GPU-Z	0%
GPU Load (%) - Highest - GPU-Z	0%
GPU Temperature (C) - Average - GPU-Z	29 degrees Celsius
GPU Temperature (C) - Highest - GPU-Z	29 degrees Celsius
GPU Power Consumption (W) - Average - GPU-Z	4 W
GPU Power Consumption (W) - Highest - GPU-Z	6 W
RAM Used by App (MB) - Current - Task Manager	-
RAM Used Total (MB) - Current - GPU-Z	2603 MB
VRAM Used by App (MB) - Current - Task Manager	-
VRAM Used Dedicated Total (MB) - Current - GPU-Z	142 MB

GPU-Z 2.68.0, Sensors

Windows Task Manager. CPU Load per Core Graph

• Borderlands 2 - Remastered, "low-latency":
  • Note: Best results highlighted in bold.

Borderlands 2 - Remastered	DXVK-GPLALL 2.7.1-4	DXVK-GPLALL 2.6.8-1
CPU Load (%) - Average - Task Manager	7%	5%
CPU Load (%) on Cores that run Application - Task Manager	25%-60%, bouncing between 2 cores	10%-50%, bouncing between 2 cores
CPU Temperature (C) - Average - GPU-Z	42.9 degrees Celsius	41.9 degrees Celsius
CPU Temperature (C) - Highest - GPU-Z	48.5 degrees Celsius	48.1 degrees Celsius
GPU Load (%) - Average - GPU-Z	14%	12%
GPU Load (%) - Highest - GPU-Z	18%	18%
GPU Temperature (C) - Average - GPU-Z	33.4 degrees Celsius	33.0 degrees Celsius
GPU Temperature (C) - Highest - GPU-Z	35.0 degrees Celsius	35.0 degrees Celsius
GPU Power Consumption (W) - Average - GPU-Z	27.6 W	25.2 W
GPU Power Consumption (W) - Highest - GPU-Z	30.0 W	30.0 W
RAM Used by App (MB) - Current - Task Manager	1204 MB	1176 MB
RAM Used Total (MB) - Current - GPU-Z	4522 MB	4541 MB
VRAM Used by App (MB) - Current - Task Manager	1023 MB	887 MB
VRAM Used Dedicated Total (MB) - Current - GPU-Z	1094 MB	1178 MB

GPU-Z v 2.68.0, Sensors: - Note: GPU-Z shows current values by default.

DXVK-GPLALL 2.7.1-4	DXVK-GPLALL 2.6.8-1

Windows Task Mangager, CPU Load per Core Graph:

DXVK-GPLALL 2.7.1-4

DXVK-GPLALL 2.6.8-1

• Borderlands 2 - Remastered, "max-frame-latency":
  • Note: Best results highlighted in bold.

Borderlands 2 - Remastered	DXVK-GPLALL 2.7.1-4	DXVK-GPLALL 2.6.8-1
CPU Load (%) - Average - Task Manager	3%	2%
CPU Load (%) on Cores that run Application - Task Manager	5%-30%, bouncing between 2 cores	10%, equally distributed between 2 cores
CPU Temperature (C) - Average - GPU-Z	42.4 degrees Celsius	41.3 degrees Celsius
CPU Temperature (C) - Highest - GPU-Z	48.8 degrees Celsius	47.1 degrees Celsius
GPU Load (%) - Average - GPU-Z	13%	11%
GPU Load (%) - Highest - GPU-Z	18%	18%
GPU Temperature (C) - Average - GPU-Z	33.1 degrees Celsius	32.7 degrees Celsius
GPU Temperature (C) - Highest - GPU-Z	35.0 degrees Celsius	34.0 degrees Celsius
GPU Power Consumption (W) - Average - GPU-Z	26.5 W	24.3 W
GPU Power Consumption (W) - Highest - GPU-Z	30.0 W	30.0 W
RAM Used by App (MB) - Current - Task Manager	1263 MB	1120 MB
RAM Used Total (MB) - Current - GPU-Z	4479 MB	4258 MB
VRAM Used by App (MB) - Current - Task Manager	1020 MB	889 MB
VRAM Used Dedicated Total (MB) - Current - GPU-Z	1092 MB	1182 MB

GPU-Z v 2.68.0, Sensors: - Note: GPU-Z shows current values by default.

DXVK-GPLALL 2.7.1-4	DXVK-GPLALL 2.6.8-1

Windows Task Mangager, CPU Load per Core Graph:

DXVK-GPLALL 2.7.1-4

DXVK-GPLALL 2.6.8-1

FPS Limiter Metrics Comparison - MIN/MAX/AVG Frametime (ms), Frametime Variance (ms)

• Educational Info:

  • AVG FPS can show general performance, but frametime is what actually shows image smoothness;
    • John Carmack, the guy behind Doom and Quake graphics programming, said that frame rendering jitter was generally perceived more poorly than lower frame rate.
  • MIN/MAX Frametime - ideally, for 30 FPS must be 33.3 ms - so the closer the result to 33.3 ms - the better;
  • Frametime Variance - difference between MIN and MAX Frametime, ideally 0.0 ms - the lower the better;
  • Frametime Variance is the reason of inconsistencies (micro-stutters) that make games feel choppy.

• Borderlands 2 - Remastered, "low-latency":

  • Note: Best results highlighted in bold.

Borderlands 2 - Remastered	DXVK-GPLALL 2.7.1-4	DXVK-GPLALL 2.6.8-1
MIN Frametime (ms)	29.6 ms	29.8 ms
MAX Frametime (ms)	37.0 ms	35.9 ms
Frametime Variance (ms)	7.4 ms	6.1 ms

DXVK HUD, Frametime Graph:

DXVK-GPLALL 2.7.1-4	DXVK-GPLALL 2.6.8-1

• Borderlands 2 - Remastered, "max-frame-latency":
  • Note: Best results highlighted in bold.

Borderlands 2 - Remastered	DXVK-GPLALL 2.7.1-4	DXVK-GPLALL 2.6.8-1
MIN Frametime (ms)	31.4 ms	33.1 ms
MAX Frametime (ms)	35.2 ms	33.6 ms
Frametime Variance (ms)	3.8 ms	0.5 ms

DXVK HUD, Frametime Graph:

DXVK-GPLALL 2.7.1-4	DXVK-GPLALL 2.6.8-1

Conclusions

1. DXVK-GPLALL 2.6.8-1 shows best results in almost all Performance Metrics, except VRAM Used Dedicated Total (MB) - Current - GPU-Z. Not a surprise, because it is basically DXVK 2.7.1 + fastest to port parts of DXVK 3.0.2 + new GPLALL improvements.
   • Differences in VRAM Used Dedicated Total (MB) - Current - GPU-Z are present, probably, because DXVK-GPLALL 2.7.1-4 uses VK_EXT_descriptor_buffer, while DXVK-GPLALL 2.6.8-1 does not use it.
2. DXVK-GPLALL 2.6.8-1 in "max-frame-latency" mode shows better results, than DXVK-GPLALL 2.6.8-1 in "low-latency" mode in almost all Performance Metrics (even RAM Used Total (MB) - Current - GPU-Z 0_0). That means that:
   • max-frame-latency mode is more performant than low-latency mode;
   • max-frame-latency mode is more efficient at the same FPS than low-latency mode;
3. DXVK-GPLALL 2.6.8-1 in "max-frame-latency" mode shows best results in all FPS Limiter Metrics and best Frametime Graph. That is expected, due to the fact that low-latency mode FPS limiter creates a higher CPU Load, which is defined by its spin-lock design.
4. DXVK-GPLALL 2.6.8-1 in "max-frame-latency" mode shows better results and better Frametime Graph, than DXVK-GPLALL 2.7.1-4 in "max-frame-latency" mode in all FPS Limiter Metrics. Again, not a surprise, but shows what a small improvement to upstream DXVK FPS Limiter can do.
5. Most of the improvements in both upstream DXVK and DXVK-GPLALL affect CPU-bound performance and RAM/VRAM usage.

Based on these benchmarks and conclusions, in DXVK-GPLALL 2.6.8-1 default dxvk.framePace mode was changed from dxvk.framePace = "low-latency" to dxvk.framePace = "max-frame-latency".