BUG: Starlight Model crippling render times on AMD RX 7900 XTX (4.5 hours for 36s clip)

Hi everyone,

I wanted to share a severe performance bug I’ve isolated regarding the Starlight Precise model and AMD RDNA3 GPUs, just in case anyone else is pulling their hair out over incredibly slow render times.

I am currently working on restoring some S-Video VHS captures, and the Starlight model is causing a complete pipeline collapse on my system, effectively abandoning the GPU.

My System Specs:

GPU: AMD Radeon RX 7900 XTX
CPU: AMD Ryzen 7 3800XT
RAM: 32gb
OS: Windows 11
Topaz Version: 1.6.1
The Issue:
When attempting to render a simple 36-second clip using Starlight Precise (Output: 1920x1080, H.265 AMD AMF), the estimated render time was 4 hours and 34 minutes. Monitoring via HWINFO showed the GPU clock speeds dropping to idle (15 MHz), indicating the GPU was either being starved of data or the software was failing to utilize it.

Chronological Troubleshooting & Isolation:
To figure out what was causing the bottleneck, I went step-by-step:

Thermals & Power: Checked HWINFO. GPU Hot Spot was well under 88°C. No thermal throttling or power limit issues.
Memory/Paging: Dropped Topaz Max Memory Usage to 20% and set Max Processes to 1 to rule out SSD paging or memory leaks. The 4.5-hour timer remained.
Model Isolation (The Fix): I kept all settings identical (Progressive, no frame interpolation) but swapped the enhancement model from Starlight to Proteus. The render time instantly dropped from 4.5 hours down to 1 minute, and the GPU engaged perfectly.
Confirmation: I swapped back to Starlight Precise as the sole enhancement model. The render time instantly jumped back to 4.5 hours.
Conclusion & Workaround:

My hardware, memory allocation, and power delivery are functioning perfectly (as proven by the 1-minute render with Proteus). The issue is strictly isolated to the Starlight model failing to communicate properly with the RX 7900 XTX.

For now, my workaround is strictly using Proteus, Iris, or Dione, which all run flawlessly and utilize the AMD hardware correctly.

I have submitted an official support ticket, but I wanted to post this here to help anyone else who might be experiencing this. Is anyone else with a 7000-series AMD card experiencing this exact bug with Starlight?

The support team got your ticket Connor and the information has been shared to the devs for review of the situation. We will leave this here for other users to report if they are running into similar issues with the same hardware.

I’m experiencing the same issue.

When using other enhancement models the card runs at 100 % Utilization, ~2800-3000 MHz and ~450 Watts.

With Starlight Precise it’s only ~60 % Utilization, ~2300 MHz and 200-250 Watts.
It feels like the GPU gets barely used by this enhancement model.

What kind of baseline performance do you normally see with Starlight when nothing is bugged?

Based on what you described — especially the part where switching from Starlight to a faster model like Proteus fixes the slowdown — it’s possible you’re not dealing with a bug at all.

For me, both Starlight Precise and Starlight Mini have always had extremely long render times. That’s been normal behavior, not an error.

I wouldn’t personally use a workaround like swapping to another model, because Starlight really does operate in its own category. The heavy render time seems to be the cost of doing something that no other model can match.

So if your only symptom is “Starlight is slow,” that lines up with my experience — not a malfunction. The frequency dips you mentioned (dropping to ~15 MHz) also match what I see. My shader clocks will briefly fall that low during certain Starlight passes, and GPU utilization jumps up and down as well.

As for the wattage numbers the other user reported, that’s the only part that doesn’t line up with my setup. My GPU power may occasionally dip to around ~70 watts, but that’s rare — most of the time HWInfo shows a steady 450–550 watts throughout the entire render, even when the frequency dips occur.

Screenshot of what Hwinfo looks like on after about 1 hour or rendering Starlight Precise 2.5 on Topaz 1.6.1: