So is not having 32gb of RAM another thing the system check will pass your system despite not having it meaning the software won’t work? What’s next? As I can’t rely on that system check, could someone kindly provide a definitive list of absolute minimum requirements that If not met, means the software will not work. What about CPU? Is there a minimum requirement for that? I have a Ryzen 9 3900X that I could use to build a new system, is that ok?
I have two PCs on which I can run Topaz. My main PC (on which I am currently running Topaz Video 1.3) is pretty beefy:
-
ASUS ROG Maximus Hero z790
-
Intel i9 14900KF
-
Gigabyte RTX 5080 OC 16GB GDDR7 VRAM (compute level 12)
-
64 GB Corsair Dominator Platinum DDR5 6400MHz System RAM
The other system was put together just for Topaz work and consists of the following:
-
ASUS ROG Maximus Hero XI (PCIE v3)
-
Intel i9 9900KF
-
Asus RTX 5070Ti 16GDDR7 VRAM (compute level 12)
-
32 GB Corsair Vengeance Pro SL DDR4 3600MHz System RAM
Both systems run Topaz just fine (except for the long outstanding problem of SL Sharp and Fast not working since the Neuroserver appeared). The 5070Ti is about 10% slower than the 5080 for Starlight processing, but only slightly (at least until 1.4.0 - haven’t benchmarked 1.4 or later since I can’t run all the models). In the release announcement for release 1.6 the SL Precise 2.5 requirements are stated as:
Nvidia Requirements:
Requires an absolute minimum of 12GB VRAM with 32GB CPU RAM, though 16GB VRAM is recommended.
Targeted for GPU compute compatibility of 8.9. This is also the recommendation.
- Please note, though there might be some success with a CC of 8.6 cards, this is not
This model is included locally, using neuroserver.
AMD Requirements:
- Requires an absolute minimum of 12GB VRAM with 32GB RAM, though 16GB VRAM is recommended.
Mac Requirements:
- Requires an absolute minimum of 16GB though 24GB is recommended.
All Apple Silicon should be supported.
This is the first time I can recall a minimum system RAM requirement being provided.
Actually, all the minimum system requirements can also be found in the online docs where the minimum for all but the diffusion models is 6GB VRAM and 16GB of system RAM, while all the Starlight (diffusion based) models require a minimum of 12GB VRAM and 32GB system RAM.
Dianac, you’re such a wealth of knowledge! I really appreciate it! this is is super helpful!.
The minimum spec for AMD cpu is 2020 onwards, i think thats because of the AVX requirements. I have an Ryzen 9 3900X released in Mid 2019 but as this was a flagship CPU, it has AVX2 support and is more powerful than most CPU’s released in 2020.
It’s such a gamble to build a system around this CPU only to find it still doesn’t run.
Does anyone know of any real world performance differences between NVIDIA GPU’s vs AMD GPUs when using starlight 2.5?
Getting a cheap AMD GPU with 16GB or more ram is cheaper than getting a NVIDIA GPU with a cuda compute of 12.
Many thanks again 
AMD, like Apple Silicon, will be much slower.
A RTX 5070 Ti seems to be a bit of a sweet spot regarding cost/performance ratio.
Glad I could help. I think the requirement regarding CPUs is indeed AVX2 support. Perhaps @kyle.topazlabs could comment about Ryzen 9 3900X support.
Edited to add:
The 3900X is a Zen 2 architecture CPU. While AMD has supported AVX2 since the last of the FX series (pre-Ryzen circa 2015), the AMD post 2020 requirement looks to be based not just on AVX2, but also some other aspect of Zen 3 architecture. I’m not intimately familiar with Ryzen CPUs since the last AMD based PC I had was FX-9370 based. I’ve been an Intel girl ever since.
Hey,
Thank you for the update, you’ve been a fantastic help :). So i’ll need new RAM, new CPU and new GPU. I think i’ll graciously bow out at this stage. The forever moving goal posts of specs and costs to build a PC to run SL2.5 is just too much. I just wish the SL system checker would have told me this when before I upgraded my subscription.
Yeah, it is a terrible time to buy PC hardware. I paid less than $200 each for 32GB DDR5 kits last year. Those same kits are now ~$600. The memory shortage has kept GPU cards at or above MSRP. It is a crazy time when a GPU can cost more than a mortgage payment.
CPU wise, the AVX2 requirement was removed in Video AI back in version 4.1.0 and was a recommendation but not a full requirement any longer. Not having that instruction set could lead to slower performance with some models. The requirement is to have AVX though, as long as the CPU has that you are good from that end.
Sadly, the current state of the world is making it hard to upgrade or purchase new gear to handle the larger models. The dev and research teams are trying to find ways to scale back hardware requirements without sacrificing the quality, which has been the rough part with the diffusion-based models.
The old adage of you can get something fast, cheap, or good, but you can’t have all three, really hits with the AI models right now.
We started with the Research release of Starlight, which could not be run locally, and produced amazing results, then scaled it down to run locally with less quality. The other variants are all attempts at fine-tuning results in different directions from my understanding, and they are working on more.
That’s all true and accepted. Each of the Starlight variants do a slightly different job and once you get used to what each one can do, it is relatively easy to choose the best one for your particular need.
That said, I am still forced to run Topaz Video 1.3, meaning I have no access to SL Precise, precisely (no pun intended) because my source material needs Starlight Sharp. I can not run Starlight Sharp in 1.4, 1.5 or 1.6 since it never outputs anything, including on a fresh Windows 11 25H2 install. Starlight Fast also does not run, throwing an inference error as soon as model load completes. It has been this way now for THREE releases. Has there been ANY progess in solving this problem?
Edited to add…
I have already supplied logs and Direct-X profile,
Understood, and apologies for the delay in response. Nice use of a pun, though, even if unintended.
Trying to get the devs to dive deeper into these situations and reviewing all the logs we have collected. Had a brief conversation earlier about this and stressing to the team the frustration.
We appreciate all you do, Kyle. You’re a gem!
As it’s hard to people to get the right equipment to run Starlight 2.5 and the sheer amount of bugs, is there not an option for the cloud processing video length cap to be removed and for the price to be reduced? To upscale a whole movie would cost close to the price of a GPU to use your cloud service. $10 to render a 30 minute TV episode is reasonable.
You can do it by renting a windows server 2022 with rtx 4090 on leadergpu, it costs about $20 for 24 hours, and it will be enough for 200% upscale of a one hour long 640x480 24fps source with SLP 2.5. (The upscaling speed on rtx 4090 is about 1fps). The only issue you need to install nvidia drivers first, then Microsoft Visual C++ 2015-2022 Redistributable Package, then Topaz Video.
$10 probably wouldn’t cover the electricity to run the servers for an hour.
Thank you. As Dianac said, I think it’s a system RAM shortage.
Here’s some additional information:
Originally, before Starlight, I was using a PC with an RTX 2060 Super (which could barely run Starlight Mini, and took 4 hours to upscale a 1-minute SD video at 30FPS to HD with 3x SLM). I tried putting an RTX 5070 in it, and surprisingly, Starlight Precise passed the check, downloaded, and ran properly! So I mistakenly thought that even with 16GB of RAM, it would run, albeit slowly. Of course, it took 11 hours to upscale a 7-second SD video at 50FPS to HD at 2x SLP, so it wasn’t practical, but it processed it properly. With a 5090, the GPU processing is faster, so I thought that even though the processing time would increase with 16GB of RAM, the processing itself would still be done. Anyway, to save money, I bought a PC with 16GB of RAM and a 5090, but after reinstalling Windows, the image was still corrupted output, so I’ll just go ahead and upgrade to 32GB.
Ryzen 5 3600
16GB RAM
RTX2060 Super (8GB) → RTX5070 (12GB)
Windows 10 Home
(Google Translate)
Thank you mate, let me know want your render time is once you install more RAM please, would be very helpful to know real world expectations
When I ran a 200% 640x480 upscale with Starlight Precise on a system with 64gb RAM there was around 43gb RAM used in Task Manager, so if you upgrade get at least 48gb RAM. The speed is aroud 1fps on RTX 4090 with 24gb VRAM.
You have less than 32 GB RAM, right?
