Hello,
I am experiencing a repeatable performance problem with Starlight Precise 2.6. Other Topaz enhancement models work normally on the same system; this problem only occurs with Starlight Precise 2.6.
System specifications
-
GPU: NVIDIA GeForce RTX 5080, 16GB VRAM
-
CPU: Intel Core Ultra 7 270K Plus
-
RAM: 96GB DDR5
-
OS: Windows 11 25H2, build 26200
-
NVIDIA driver: 610.88
-
Topaz Video: 1.7.1
-
Two 4K displays
-
Intel integrated graphics also available
Test settings
-
Source: 9-second H.264 High Profile video
-
Resolution: 1280Ă—720
-
Frame rate: approximately 24 FPS
-
Starlight Precise 2.6
-
Sharpness/Softness level: 1
-
Output: 2560Ă—1440, 2Ă— upscale
-
No frame interpolation or additional enhancement models
-
The complete source file is exported; no In/Out trim is being used
What happens
At the beginning of the export, the RTX 5080 appears to work normally. GPU power consumption remains above 250W and processing progresses for approximately two or three minutes.
After that, performance suddenly collapses:
-
GPU power drops from over 250W to approximately 84W and never recovers.
-
Processing speed drops to 0.0 FPS.
-
The estimated remaining time increases to more than 21 hours for a nine-second video.
-
GPU temperature is only around 46°C, so this does not appear to be thermal throttling.
-
GPU clocks remain high, but GPU memory-controller utilization drops to approximately 3%.
-
Neuroserver begins using a very large amount of shared system memory.
During the stalled state, I measured approximately:
-
10.8–11.1GB dedicated GPU memory used by
neuroserver.exe -
14.6GB shared GPU/system memory
-
Approximately 23GB working-set RAM
-
Approximately 33GB private committed memory
The GPU may still be reported as highly utilized, but power consumption and actual processing throughput are extremely low. This looks like VRAM oversubscription followed by NVIDIA/Windows system-memory fallback and severe memory-transfer thrashing.
A shorter five-second clip has completed successfully before, but nine-second and longer clips repeatedly enter this stalled state. A 30-second 720p clip can produce an estimated processing time of more than one day.
Troubleshooting already performed
-
Completely reinstalled Topaz Video
-
Restarted Topaz Video and Windows between tests
-
Tested the Topaz maximum-memory preference at both 70% and 100%
-
Confirmed that no other GPU-heavy applications were running
-
Confirmed that temperatures, GPU clocks, PCIe link and power delivery were normal
-
Confirmed that other Topaz models work normally
Changing the maximum-memory preference did not appear to constrain Starlight Precise 2.6. In the logs, Neuroserver was still launched with:
--max-gpu-mem 16
This occurred even when the application preference was set to 70%.
Questions
-
Is this a known memory leak or VRAM-oversubscription issue in Starlight Precise 2.6 or Neuroserver?
-
Why does the application continue to pass
--max-gpu-mem 16when the maximum-memory preference is reduced? -
Is there a supported way to force a smaller temporal batch, tiled processing, or a lower effective GPU-memory budget such as 12GB?
-
Is 16GB VRAM genuinely expected to process 720p → 1440p footage with Precise 2.6 without falling back to shared system memory?
-
Are there any additional logs or diagnostic files I can provide to help reproduce this?
The official requirements recommend 16GB VRAM for Starlight Precise 2.6, so I would expect this 720p test to remain usable on an RTX 5080.
Thank you.
