Starlight Precise 2.5 working perfectly after a week of repairs / tweaks

Topaz Video General
1

Greetings fellow renderers,

I’m using Version 1.6 with Starlight Precise 2.5 and it is working without any issues at all after about ten renders. The one that is running now is 7.5 hours long with 1 more hour to go. 100% success so far. I’m 3X upscaling 720X480 ProRes 422 files to 2160 X 1440 resolution.

A quick rundown of my week of upgrades, noob mistakes, a virus and system tweaks incase this helps anyone

Updated my CPU from a stone age model to a bronze age model - system overheated

  • installed a larger stock CPU cooler - system overheated
  • installed a new Scythe Fuma 3 air cooler - system overheated
  • removed the protective plastic from the Fuma 3 cooler ( a common mistake) - system good

Installed 3 extra 120 mm Arctic Pro fans that are externally powered by a 12V 2.5 A power supply that has a switch with and two fan power leads. Added a Noctua NA FC1 fan controller which plugs into the power supply and controls the fan speeds with a dial. With more fans they don’t have to run as fast and I have dialed in the speed that keeps the CPU at about a 74 C average temperature with the highest spike to 90.5 C, which is very safe. If you get the Noctua NA FC1 it detaches from the power supply too easily. I used small wire ties to secure the connection.

Next my keyboard got totally erratic so I plugged in another one and it had the same problem. Then Windows Defender alerted me that it had isolated a virus that had attached itself to Open Hardware Monitor named the WinNT/ Winring0 virus. Cleaned and rebooted and back to normal. Now I’m using HWiNFO with just the CPU sensor status enabled.

Then my computer completely lost power. My power supply was about 18 years old and the parts dry out and start to fail apparently. Installed a new 850 Watt power supply and my system ran noticeably better and the CPU temps dropped a bit.

Then Topaz Starlight shut down during a render.
I read “Reports around Starlight Precise 2.5 indicate a neuroserver memory leak / out‑of‑memory behavior, especially with certain containers (MKV) and when using trims. Users have seen it eat essentially all system RAM (e.g., 64 GB) before failing, causing crashes or stalls,” I’m not sure if this information is accurate or not.

My error code
Name: System Source: nvlddmkm “That nvlddmkm Event ID 153 means the NVIDIA driver/GPU crashed”

I was advised “Use DDU in Safe Mode to remove all NVIDIA drivers, then install the latest NVIDIA Studio driver directly from NVIDIA’s site.” I downloaded and installed DDU and booted into safe mode. DDU was on my screen in safe mode and I ran it to clean out the old drivers and after boot immediately installed new STUDIO drivers using “Custom Install” Clean Install and unchecked the audio option.

Next I went into the BIOS and changed my GPU’s PCI slot option from automatic to Gen 4. I had previously set my CPU power to Eco mode at 95 watts to back off on the power a little.

I reboot Topaz Video each time and reboot my computer regularly, both of which may be unnecessary but it’s working. I’d rather restart and reboot rather than lose a long render.

MSI Ventus 3 GeForce RTX 5070 Ti, AMD Ryzen 7 5800XT, MSI B550 PRO-VDH mobo, 32 Gb 3200 RAM

Good luck everyone! Dave Jeffery

2

Thanks for posting that very detailed account. I found it very educating. I’m a Mac guy who recently took a windows plunge and got my first pc - a Lenovo legion with an RTX 5090 card so everything related is new to me. There seems to be so many variables and settings I am completely unaware of so I appreciate hearing about challenges related to GFX cards and software that others have come across and fixed on their systems. It helps me get a fuller understanding of the somewhat complicated windows nvidia topaz world.

3

That’s a nice PC built experience diary :slight_smile:

Tip for anyone who still has some space left in their case. I’m a big fan now slot fans, which tend to occupy a niche role in the PC-building world—unjustly so.

front-to-back airflow is impaired by top and bottom fans/airflow, this is not ideal. That’s why top-mounted fans often provide only limited benefits, and even a push-pull setup at the top is hardly any better.

In almost every case, more air is being pushed in through the front than is being exhausted out the back of the PC—that’s the real issue. Slot fans help solve this by pulling more air full trough the case and extracting GPU’s hot air directly, which also means the CPU ends up drawing in cooler air.

Slot Rear Exhaust Cooling Fan
Slot Rear Exhaust Cooling Fan543×211 11.3 KB

For quiet operation, I’d recommend running these fans at 7V instead of 12V, by changing the pins or use a fan controller…sure you can install more than one of them.

4

Oh no! I’m so glad you figured that one out.

5

The old AMD CPUs would get instantly grilled if you forgot to remove the protective film from the cooler. Those were the days :grinning_face_with_smiling_eyes:

6

I use one of those in my file server to keep air flowing around the Adaptec 16 channel RAID controller. The RAID chip gets ridiculously hot (85 degrees centigrade is considered “normal”) without the exhaust fan.

In my Topaz machine I have a dual chamber case, so the PSU is on one side and the motherboard on the other. So I use three intake fans on the bottom, another three intake on the side, three exhaust fans on the top blowing through the CPU radiator, and one exhaust on the back. I track ambient temperature inside the case (all the fans are indexed to this sensor except for the radiator fans which index off of the CPU coolant temp) and it never goes over 28 C in there. CPU and GPU temps hover around 60 C under load.

7

Hello people of topaz ! I don’t know if it is of any use in your specific case, but I solved all my GPU temperature issues when using heavy and long SLM/SLP workflows by installing MSI Afterburner and tweaking the GPU temperature curve manually. Works like a charm. And you can set the app to launch with Windows, it consumes very little resources and that way, you don’t forget starting it when you run a heavy workflow.

8

Can you provide the settings you are using?

9

Capture d’écran 2026-06-10 165621
Capture d’écran 2026-06-10 165621504×848 36.5 KB

Here you are, but I have a GPU which runs hotter than consumer GPUs, it’a an A6000pro blackwell, it gets hot very fast but it is meant to run hours at 75-80°, so adapt the settings to your GPU of course.

10

Thank you. I have a 5090.

So your only tweak is temperature. No clock or voltage…

11

You are welcome. Yes, no clock or voltage as I use heavy comfyUI workflows where undervolting will slow down inference, I already tried. Sure, AI doesn’t pump your GPU’s clock or power consumption as much as gaming, but in the end, when a workflow runs for 15+ minutes, undervolting especially WILL slow it down.

12

you can check older Threat here

I’m super happy with my 850mV limited RTX5090. Then you can push Vram clock at least +200Mhz (value +2000 into Afterburner)…later with the newer Afterburner Beta version, i set ram clock into Afterburner to the max value of +3000…this worked stable, bust just for a while, then i must lower it to about +2700 to get a stable card.

With the 850mV limited card and increased Vram clock, you just loose about 5%-7% speed compared to the factory default card, TDP is about 200W lower, what i see is, the card takes max 400W and this are peaks!

13

Thanks for the link. :grinning_face_with_smiling_eyes:

14

5% to 7% slower, and around 400 watts…honestly sounds like you turned your 5090 into a 5080. That is just about exactly the performance I get from my 5080.

15

ASUS ROG Crosshair X870E APEX
Gigabyte RTX 5070 Ti Windforce OC SFF 16G
AMD Ryzen 9 9950X
ARCTIC Liquid Freezer III 360
RAM (G.SKILL Trident Z5 Neo 64GB)
ASUS ROG Strix 1000P Platinum ATX 3.1
COOLER MASTER MASTERBOX TD500 V2 MESH

My fps on starlight precise 2.5 drop down to 0 after a while, suddenly says for a 4 min video 1-2 days for process. Do you have stable fps?

16

It looks a lot like VRAM oversubscription, which is causing it to use slower system RAM. The question is why. Make sure other programs aren’t using up a lot of VRAM. With your graphics card, if you take an SD (720x576) video and use Starlight Precise 2.5 for a 3x upscale, the processing time for a 4-minute video should be around 6–7 hours.

17

A 5–7% FPS reduction for 30% lower power consumption is a good trade-off for me. The TDP of the RTX 5090 is 575W, but mine now runs at 400W or less, with 400W being the peak power draw, overall it is about 350W.

The RTX 5080 is about 15% slower factory comparisons, while consuming roughly the same amount of power as my undervolted RTX 5090. So yes, you could argue that the 5080 offers better value for the money, but it’s still slower than my undervolted 5090 and has less VRAM.

I bought my RTX 5090 before prices went crazy. If I were buying today, I probably wouldn’t make the same choice. But undervolting it was the best decision I could have made, and I have absolutely no regrets about it.