That canāt be repeated often enough, especially for upscaling (not so much slomo).
I will post benches later when some processing of 3.5.0 has finished but I did just try using 2 instances of TVAI on a 1080p 2x upscale, Artemis High, 50p video. rtx 4080 with 5800x CPU and DDR4, 32GB.
Result:
one instance, GPU avg 28%, CPU 100%. 5.2 fps.
two instances,GPU avg 28%, CPU 100%, 2.5 to 2.6 fps.
Now THATās a CPU block for ya! So much talk about powerful GPUs but really, for upscaling, people should think more about their CPU and matching memory, even those known to be fast in their day.
This seems to be true mostly for Windows systems.
On the Mac you donāt get that high CPU load - which is fine as that way you can fully use the computer for ānormal tasksā without noticing that thereās an encode running in the background.
I just discovered this benchmark comparison and donāt have the numbers for my 3080 Ti, but I know the wait times have been the same after upgrading to a 4090.
Donāt understand what the problem is.
A CPU bottleneck, really? My numbers are slightly worse than a 7950X and only the 13900K looks to be an improvement, based on the limited examples here.
Regardless, it feels like there shouldāve been quite the difference in wait times going from a 3080 Ti to a 4090, not more or less the sameā¦ which got extended by 1 hour with the April or May update for some reason, by the way.
Itās starting to look like bad optimization for Nvidia cards or at least for the 30 & 40 series, especially since that update. Maybe others can confirm having their wait times increase by 1 hour with Artemis.
Bottlenecking aside, note the fps difference between my real world 2x upscale of Artemis and the benchmarks. 5.2 fps real world, 12.31 fps benchmark.
When I upscale just a few frames however, the āOutputsā report 5.2 fps for 10 minutes of 1080p and also 25 frames; 7 fps for 12 frames, 8.6 for 8 frames andā¦ 15 fps for 5 frames.
If thatās true, it means TVAI isnāt using all available Tensor & RT cores and is indeed an optimization problem.
The fact that the CPU is in 100% utilization while GPUs are barely tapped for their potential shouldāve been the 1st sign, but I saw a lot of people say itās normal so didnāt question it.
3080 Ti vs 4090
Tensor Cores 320 vs Tensor Cores 512
RT Cores 80 vs RT Cores 128
Hi everyone. I ask for information as a TVAI user on Mac. I would like to switch to Mac Silicon M2. Do you recommend the Macbook Pro M2 Max (12/38 Core) or Mac Studio M2 Ultra (24/76 Core).
Looking at your tests I didnāt find much difference in performance
It would be possible to see the processing differences in fps for extreme encoding
With very extreme encodings the speed differences do get smaller for all hardware.
When doing 4K=>8k up scaling even a NV 4090 is only somewhat faster than other GPUs as in such scenarios RAM and HD speed as well as CPU do come increasingly more into account.
So youāll benefit from a really fast GPU the most in SD/HD sources and with 2x upscales.
See the benchmarks.
