If you currently own and use either of these cards, what did you upgrade from? I would be interested to know your real world gains, and for your individual use case.
Please share!
I realize the beta now supports multi gpu, so I suppose I’m only referring to those with single gpus. But if you have experience already with multi gpu etc also please feel free to share your real world gains from single card maybe too!
Upgraded from 4090 and it’s slower 
This is actually common for software that aren’t optimized for newer GPU generations.
The Blackwell 6000 Blackwell is approximately 10-15% faster than the RTX 5090 when it comes to video upscaling.
The Blackwell has 24,064 CUDA Cores and the 5090 21,760. The Blackwell has 96 GB of GDDR7 memory, and the 5090 1/3 that amount 32 GB GDDR7.
My challenge is I can’t translate these specs into real world performance. I would like to close the gap between paper specs and actual results.
Blackwell vs Lovelance:
Blackwell should be able to handle more work in Parallel, you will notice this when running more than one upscaler at the same time.
Blackwell Pro has the MIC feature, but i think its linux only.
Hello TPX,
I am rooting for continued improvement with the Series-50 Blackwell instruction set.
Yes and No.
More memory for AI means, you can load bigger models.
More memory means too that you can use more software in Parallel.
And more memory means higher speed if you run out of memory with your software.
Blackwell has a RISC-V chip only for Sheduling, so it should be able to handle more work in parallel, thats also true for the chip design, backwell is a full parallel design compared to Lovelance.
Blackwell is not faster because of compute power but because of Parallel power.
You should pair a blackwell GPU with a big Cache CPU to handle Multitasking better.
The Blackwell instructions (fp4) are only good for LLM (ChatGPT) and not for image creation and editing.
Thank you, Thomas, for the insight you have given me.
“More memory for AI means, you can load bigger models” like: Recover V2, Focus V1, etc.
Caz,
I have this…take with a grain of salt.
Hardware Performance Hierarchy:
Hello Caz,
*I think the good news is you have leads to 2 Topaz Video AI users you can contact: *
rfrankway - who uses a 6000 Blackwell and a 5090 seemingly at the same time, and
cskzhi - who uses a 5090.
The problem with all the ‘benchmarks’ even Puget systems’ is none of them actually show a real-world test. I think it would be easy for one of these reviewers to have three 5 or even 10-minute clips and use different models on each, some maybe run twice like how some do SLm + Proteus or Nyx + Proteus or Adding Iris etc.
The models aren’t important, what is, is a REAL-WORLD test. I made another similar thread a long while ago.
I’m less worried about SLm and more concerned with how for example how a 480p>1080p using mostly Proteus would go if a user went from maybe a 3090 to a 5090. Having a graph including 5-8 cards. Obviously AMD is not as supported but I think reviewing a 9070 xt/7900 xtx along with a 4090, 5070Ti, 2080Ti, and maybe something like a 3060/4060 for reference.
This is not to say there are ZERO results, I have seen a few youtube reviewers and others do actual real world testing and it was fascinating to see!!! None come to mind off the top of my head but I’ll update this thread if I find them, even if the tests are years old.
ForSerious made a thread a while ago dedicated to this topic of hardware upgrades. Well worth seeing even given its 2+ years age
Thought provoking! Thank you, cas.
Thank you, cas, for hightlighting what I also think is an important issue: “The problem with all the ‘benchmarks’ even Puget systems’ is none of them actually show a real-world test.”
Even two and a half years I found What hardware increases speed in TVAI to be instructive.
