+- Selur's Little Message Board (https://forum.selur.net)
+-- Forum: Hybrid - Support (https://forum.selur.net/forum-1.html)
+--- Forum: Problems & Questions (https://forum.selur.net/forum-3.html)
+--- Thread: Torch Addon ... (/thread-3510.html)
RE: Torch Addon ... - ToiletDuck - 20.12.2023
(20.12.2023, 22:16)Selur Wrote:Quote:but is it now ? Have you add up the time it took to prepare before the actuall encode started ?The .engine file is build for different settings, so you can reuse them for different files as long as the settings stay the same,...
In my case it was named:
Very HELPFULL INFO right there ↑ , much obliged Sir Selur
But then again.. ↓
(20.12.2023, 22:16)Selur Wrote:(vs-mlrt uses .engine as extension)realesr-general-x4v3.pth_NVIDIA GeForce RTX 4080_trt-8.6.1_720x576_fp16_workspace-1073741824_denoise-0.5.pt
Using TensorRT doesn't make sense if you use short clip, but it does make sense for stuff that is longer.
Cu Selur
Are you saying... that the .enginge file created is of no use to longer clips
? You've lost me right there ! Are you talking about clips / media files of mins vs hours ? If so , why is that ?
cheers,
td
RE: Torch Addon ... - Selur - 20.12.2023
No.
for short clips: using TensorRT doesn't make sense if you need to build the engine file (if you already have one, it does make sense)
for longer clips: it makes sense to use TensorRT even when the engine file needs to be created.
Usually I already have an engine file, since it did get created for preview if it didn't exist.
Cu Selur
RE: Torch Addon ... - ToiletDuck - 20.12.2023
(20.12.2023, 22:28)Selur Wrote: No.
for short clips: using TensorRT doesn't make sense if you need to build the engine file (if you already have one, it does make sense)
for longer clips: it makes sense to use TensorRT even when the engine file needs to be created.
Usually I already have an engine file, since it did get created for preview if it didn't exist.
Cu Selur
So... if i read you correct... building an engine file doesn't make sense for smaller clips, because .. small performance gainz. ..then ?
Cuz, it sure looks like it over here ...!
With Tensor i gained like only 1-2fps .. that all.. i just get an lil friendly gesture
NOw.. tensor + fp16 i gain like 4 - 6 fps
... tested on small samples so far... hence my question earlier..EDIT: atm, speeds gain went up from 3fps to 9fps
... how delightfull ..So yeas, half of cuda / RT cores = half of your speed Selure... simply put..
cheers
TD
RE: Torch Addon ... - Selur - 20.12.2023
If building the engine + encoding with the faster speed takes longer than encoding with the slow speed, it does not make sense.
So for short clips it does not make sense to use TensorRT if you have to build the engine file.
Quote:So yeas, half of cuda cores = half of your speed Selure... simply put..TensorRT uses the tensor cores.
Cu Selur
RE: Torch Addon ... - ToiletDuck - 20.12.2023
(20.12.2023, 22:38)Selur Wrote: If building the engine + encoding with the faster speed takes longer than encoding with the slow speed, it does not make sense.
So for short clips it does not make sense to use TensorRT if you have to build the engine file.
Cu Selur
Indeed, i read where you want to go with this statement ↑
And rest Asure , even on very small mini-minute samples it's an PLUS+++ to enable tensor in my case..
Prep takes like only a few secs, but encoding time decreases with litteraly minutes on a 3 minute sample ^^ ..
So go figure Sir Selur, WHAT WOULD YOU CHOOSE TO ACCEPT 0^0 ?
(20.12.2023, 22:38)Selur Wrote: TensorRT uses the tensor cores.
Cu Selur
HOld on.. wait a MInute ... WAIT A MINUTE ... STOP
!!Next to cuda and RT cores, a green card has Tensor cores aswell?
Nvidia HQ mentioned in the specs of the card that it has a Xth gen T-core?
But it doesn't mentioned how many , where is that "officialy" announced ? Not talking about 3rd party info..
cheers,
TD
TD
RE: Torch Addon ... - Selur - 20.12.2023
https://www.quora.com/At-an-architectural-level-how-are-tensor-RT-and-CUDA-cores-different-What-makes-tensor-and-RT-core-so-much-better-at-DLSS-and-ray-tracing has a good explanation (by Heikki Kultala) about what tensor, rt and cuda cores are.
Geforce RTX 4080 has 304 tensor cores
see: https://www.techpowerup.com/gpu-specs/geforce-rtx-4080.c3888
Geforce RTC 4060 Ti has 136 tensor cores.
see: https://www.techpowerup.com/gpu-specs/geforce-rtx-4060-ti-16-gb.c4155
(the 4090 has 544 tensor cores)
alternatively, see https://en.wikipedia.org/wiki/GeForce_40_series under 'Core config'.
Side note: usually the relative amount of cores are similar so if is card has 5 times more cuda cores than another card, it usually also has 5 times more tensor and rt cores than the other card. So a rough estimation can usually be done by just looking at the cuda core counts of the cards.
The generation of the cores defines how powerful they are. (newer generation cores can compute more than older generation cores)
Cu Selur
Ps.: going to bed now
PPs.: NVIDIA usually has such infos in their white papers, see: https://images.nvidia.com/aem-dam/Solutions/Data-Center/l4/nvidia-ada-gpu-architecture-whitepaper-v2.1.pdf (Tensor core count and generation usually is not that interesting for gamers, to it's not listed on the main pages,...)
RE: Torch Addon ... - ToiletDuck - 20.12.2023
(20.12.2023, 23:17)Selur Wrote: https://www.quora.com/At-an-architectural-level-how-are-tensor-RT-and-CUDA-cores-different-What-makes-tensor-and-RT-core-so-much-better-at-DLSS-and-ray-tracing has a good explanation (by Heikki Kultala) about what tensor, rt and cuda cores are.
Geforce RTX 4080 has 304 tensor cores
see: https://www.techpowerup.com/gpu-specs/geforce-rtx-4080.c3888
Geforce RTC 4060 Ti has 136 tensor cores.
see: https://www.techpowerup.com/gpu-specs/geforce-rtx-4060-ti-16-gb.c4155
(the 4090 has 544 tensor cores)
alternatively, see https://en.wikipedia.org/wiki/GeForce_40_series under 'Core config'.
Side note: usually the relative amount of cores are similar so if is card has 5 times more cuda cores than another card, it usually also has 5 times more tensor and rt cores than the other card. So a rough estimation can usually be done by just looking at the cuda core counts of the cards.
The generation of the cores defines how powerful they are. (newer generation cores can compute more than older generation cores)
Cu Selur
Ps.: going to bed now
Why doesn't Nivida publish and mentioned this kind of valuable info in their specs at their official webpage is beyond me
... I know you can get the info from all sorts of places... but .. it isn't alway's reliable.. not even wiki leaks are alway's spot on (outdated..)...
Thanks for the info anyway..
Sweet dreams..
TD out..
RE: Torch Addon ... - Selur - 20.12.2023
added link to whitepaper,..
RE: Torch Addon ... - ToiletDuck - 20.12.2023
(20.12.2023, 23:31)Selur Wrote: added link to whitepaper,..
I know it's @ wiki's...
But thats jus it.. why don't they just add all that info to their official website instead..
with all the info from different sources flooding the net nowdays, not to mention fake nonsense nowdays you know ..?
If it is published at their official site, they can't deny they were wrong, lying, or claim the info is from a 3rd party (unofficial source) and therefor false afterwards.. thats my point..
thanks for the info anyway.
cheers,
TD
RE: Torch Addon ... - Selur - 21.12.2023
Here is the link again:
https://images.nvidia.com/aem-dam/Solutions/Data-Center/l4/nvidia-ada-gpu-architecture-whitepaper-v2.1.pdf
This is officially from the NVIDIA homepage.
Cu Selur