I currently utilize a NVIDIA GeForce RTX 5060 Ti 16GB which leans heavily on the Starlight series for local rendering. I wondered about anyone’s experience with using a 5090 32GB and any experience with a steep upgrade and performance improvements. Currently I’m in the process of using the Starlight to render 2 hour digitized Hi8 videos and this takes about 4 days and I have about 75 videos to render. In researching, I have come across the Suprem Liquid version and wondered if anyone is using this to good results?
If you have any thoughts or advice on the possible upgrade, I’d appreciate hearing from you. It would be costly, but maybe worth the investment? Thanks!
Performance will be around an RTX 4090 - but rock solid with ECC enabled, its TDP is 300 watts - so 150 watts less than 4090 and 275 less than RTX 5090.
Price is a little bit higher but if you do resell it you will get more back than for an gaming card.
Thank you for this information and for your suggestion. Installing the RTX 5090 would possibly also mean an upgrade to my beefy power supply, but would be unnecessary with the RTX Pro 5000 using the lower wattage. Thus, the cost difference may be little if any. I am grateful that you took the time to respond and I’ll research further and consider!
on the 4090 i had small improvements only. CuNDD 9.24 is key to use for uplift but for the 5090 i had 30-44%+ improvement in speed with temporal clip chunk settings to 241, temporal overlap 21 and VAE Cap at -1 (so no cap). You need at least 64GB system RAM. When i save money for the 6000 pro i would expect also higher improvements. What I observed is that the launcher runs into close to 27-28 VRAM allocation stays there for 1 minute, speed is boosted 3-5x and then goes back to 15-19GB VRAM allocation at slower speed for 8-10 minutes then again higher allocation. I have not fully figured out a way to keep it at higher VRAM allocation. But overall it’s a dramatic improvement. The biggest impact though is CuNDD 9.24
My undervolted RTX 5090 has a TDP of around 400W, so the gap between the 300W RTX 5090 isn’t that huge. Both are based on the Blackwell architecture, the lower TDP of the RTX Pro 5000 comes from having fewer CUDA cores (14,080 vs. 21,760), and tensor cores.
ECC memory can lead to better stability. On the other hand, the memory bus is 384-bit, while the RTX 5090 has a 512-bit bus. More VRAM is only useful when you use it, so maybe you can benefit from it with the SLP starter tool?!
However, gpu prices are very high, personally for private use I wouldn’t buy a super expensive 5090 anymore today (i payed half price). The RTX 5080 is price-interesting, but unfortunately, the card has “only” 16GB VRAM. Another option could be the RTX 4090 with 24GB of VRAM, which you can get for about half the price of an RTX 5090…if you can find one
Thank you for this information. I undervolt my RTX 5090 to 85% or roughly 400 watts and notice minimal degradation of rendering speeds. I intend to use your suggestion and alter some of the operational characteristics of the 5090 when the current rendering schedule is done in a couple of days. I like the idea of the larger batch size as the rendering completely ceases during those data transfers. With the RTX 5090 offering 32GB of RAM and then my CPU has 64GB using a 4TB SSD this should work perfect for me.
Purchased an RTX 5090 about 3 months ago and rendering speed for a 2 hour digital camcorder video using Starlight Sharp dropped from 4 days with the RTX 5060 to just under a day with the RTX 5090.
Yes, undervolting the RTX 5090 to 850mV reduces power consumption by 25-30%. If you increasing the VRAM clock by 312 MHz (native), which is +2500 in MSI Afterburner, the performance loss is only around 7% compared with the factory-default 5090. This provides a very good performance-per-watt balance, especially for long-running rendering workloads i can recommend.
Thank you for this additional info. Without messing around with the VRAM clock, I already show an average GPU clock of 2540 shown in both Afterburner and GPU-Z. The Topaz Sharp really makes those older 1990’s videos very clean and detailed compared to the original analog Hi8 tape used from the camcorder. I’ve also applied this rendering to a few older music videos and the difference is striking.
