I use in real live the proteus model with manual setting and for my experience I obtain +/- 11 fps.It takes more or lex 5 to 6X the total output lenght of video
Iâam thinking make a upgrade to my pc and I am thinking something like this
GPU: Rtx 5080 (I donât have budget for the 5090 or 4090)
Memory: 2 X 32 Gb DD5 6000Mhz CL30
My doubt is which processor I choose, Iâam thinking between 9950x or 9900x ( the 9950 has more 4 cores and 8 threads and it costs more 140âŹ)
The purpose of this upgrade is to increase my actual fps using the same topaz video ai configuration
Main question: It is worth pay more 140⏠for 9950x or should I use this money to buy 128Gb instead of 64Gb of memory and buy the 9900x?
I only use the first 8 cores of my 9950X3D for TVAI. With most TVAI tasks, itâs faster to use only the cores from a single CCD. Even disabling hyperthreading on first 8 cores can boost performance. Therefore I think a 9700X could actually be faster than a 9900X, at least if nothing else is running in parallel (thatâs only the case when export encoding is performed on GPU, see Edit below).
Edit:
As ForSerious showed, when utilizing CPU based export format like TIFF images, the CPU utilization can be much bigger compared to a GPU based export format (eg. h265 on GPU) . On my 16 core CPU I actually measured twice the CPU utilization for TIFF export.
In this case, a 9700X with only 8 cores will most likely perform worse than a 9900X with 2x6cores. And a 9950X will probably again perform a little better than a 9900X. When using GPU based encoding, a 9700X should perform on par or sometimes faster than a 9900X in default configuration.
There are also more options in TVAI, like dynamic parameter detection which add to the CPU load and might make a 9900X with more cores faster than a 9700X.
And fast 32GB RAM will be also of more use than 128GB of slower RAM. TVAI doesnât really use more than 16GB.
Thank you for your help. When you say a fast 32GGB you mean a ram with more clock speed or with a lower CL or both? Could you give me a example of the best ram configuration?
One more question. What is the cpu from Intel that have the same perfomance of 9900x/9950x?
The hand tuned maximum u could achieve would be 6400 cl30 RAM but thatâs a lot of trial and error and sometimes the cpu (specifically the memory controller on the cpu) cant handle that speed.
But u should be able to buy a pre configured ram kit with 6200 or 6000 cl30 speed which should give u about 80-90% of that performance.
The latest intel cpu, the 285k/265k dont seem to perform very well with tvai if u look into the performance threads of the forum.
Therefore the generation before, a 14700K/KF would probably be the recommendation. It could even be faster than the Ryzen CPUs in TVAI because it supports faster RAM, but I dont have first hand experience with it.
The 9950X3D might be a special case. My 9900X uses 10 cores heavily with Proteus. There was something I was doing not long ago with TVAI that maxed out the cores and made me start wondering if I should have got the 9950X.
Just try to assign only first 6 cores to ffmpeg process in task manager while tvai is performing one of your jobs and check performance. 6 cores could be on the low side but for my 7950 and now 9950x3d, using only the first 8 cores of ccd 0 is faster in my Proteus jobs, which is why a single ccd 8 core 7900x could be faster than 12 cores of 9900x.
On my 9900X.
10 min 16:9 DVD clip with no black bars to FHD using Proteus manual.
Here are the results:
One CCD: 0:14:28.25
Half and half: 0:17:55.35
12 cores: 0:08:10.11
Unlimited: 0:05:57.37
This is including the time it took for me to limit the cores in Task Manager, so if it took me longer to set them, that would only make the time shorter since unlimited was the fastest.
Interesting. There are many variables of the settings. I neither use âRecover detailâ nor âDynamic Settingsâ. And I encode to H265 with AMD GPU (better quality compared to NVIDIA).
In my testing I found that âRecover detailâ lowers performance so much that the difference between the core settings becomes almost negligible.
âDynamic Settingsâ on the other hand seem to benefit from more cores which might explain the difference in our findings since I usually use pure manual settings which requires less cores.
For setting core affinity I use free version of Process Lasso https://bitsum.com/ .
If there is any scaling involved then yes, CPU + RAM often creates a bottleneck for the GPU since image scaling is performed by CPU in RAM. âRecover Detailâ and âDynamic Settingsâ are also CPU + RAM dependent and can create an even bigger bottleneck.
Only the Starlight model is 90% GPU since itâs so slow on the GPU so that any scaling or other tasks on the CPU are essentially never a bottleneck.
Unlimited is the default of use all cores as needed.
The fps doesnât matter really, only the time taken to complete the task. I donât trust the fps counter in TVAI and the one in ffmpeg is just an average of the whole timeâso when I paused it to set the cores it also reduced the fps number a ton. [Edit: so for the reported times, I didnât bother to pause it, since that would alter the time taken.] Iâm just using Proteus manual settings. Nothing else. I did run it with a script and output to tiff, so thereâs no GUI to bog anything down.
Tiff encoding probably takes up a good chunk of CPU, at least more than any GPU based encoding. So yes, if somebody uses CPU based output encoding, more cores are required.
Hi, I checked the CPU utilization when using default all cores with TIFF export (using CPU) and h265 export (using GPU). There is a pretty big difference on my system. TIFF export on CPU puts almost twice the load on the 16 cores: 54% vs 28% CPU utilization. Wither fewer CPU cores, the difference could be even bigger. I also found that in the case of TIFF export, using all cores perform better than a single CCD. Just like you found.
Therefore, I will annotate my recommendations in the first posts that these are only valid if the output encoding is performed on a GPU and not the CPU.
H265 export performed by AMD iGPU (total CPU utilization of 28%)
I have a 9950X3D. How can I change the settings to only use the first 8 cores for TVAI? Iâm completely new to AMD or anything computer related. haha.
Ok, itâs not a standard thing to do but can be beneficial in TVAI.
The primary possibilities are
Windows Task Manager: Details Tab in the Advanced Mode gives u a list processes and there u can right click on ffmpeg.exe while a TVAI job is running and use âSet affinityâ. Main drawback is that u need to do this every time u restart a TVAI job.
I use Process Lasso https://bitsum.com/ in the free version. There u can pre-define various affinity settings (e.g only first 8 cores, âŠ) and permanently associate processes like ffmpeg.exe with an affinity profile. The details u can find in the help or online
Whatâs the point in outputting to a lossy format? The hardware implementations of H.264 and H.265 are bigger and more lossy than the software versions. I need to output to something lossless to be able to convert them optimally into a final viewing format.
Also, before everyone gets too excited by your findings, you have not supplied the times is takes to process a longer than two minute video in all the CPU configurations.
Here u go
10:10 min DVD video 960x540 upscaled 2x to 1920x1080 with Proteus all manual settings, no recover detail, export with AMD GPU to h265
5090 downclocked to 2100Mhz, 32GB RAM 6200CL30, Win10
made a batch file with 3 exports and time output, showing the minutes:seconds here:
all cores: 05:47-09:42 = 3:55 = 235s /2.6x (as shown by TVAI command line)
ccd0 ht: 09:42-13:14 = 3:32 = 212s /2.84x (as shown by TVAI command line)
cc0 no ht: 13:14-16:43 = 3:29 = 209s /2.94x (as shown by TVAI command line)
Regarding lossy vs lossless export. Just different needs I guess. I primarily work with FHD, UHD or FHD â UHD movies. For the big difference in output size, lossless is usually not worth it for me.
export command:
ffmpeg â-hide_bannerâ â-iâ âF:/CPU/DVD 10min test.mkvâ â-sws_flagsâ âspline+accurate_rnd+full_chroma_intâ â-filter_complexâ âtvai_up=model=prob-4:scale=0:w=1920:h=1080:preblur=0.42:noise=0.23:details=0.21:halo=0.28:blur=0.31:compression=0.18:device=0:vram=0.95:instances=0â â-c:vâ âhevc_amfâ â-profile:vâ âmainâ â-profile_tierâ âmainâ â-tag:vâ âhvc1â â-pix_fmtâ âyuv420pâ â-gâ â30â â-b:vâ â24Mâ â-anâ â-map_metadataâ â0â â-map_metadata:s:vâ â0:s:vâ â-fps_mode:vâ âpassthroughâ â-mapâ â0:s?â â-c:sâ âcopyâ â-movflagsâ âfrag_keyframe+empty_moov+delay_moov+use_metadata_tags+write_colrâ â-bfâ â0â â-metadataâ âvideoai=Enhanced using prob-4; mode: manual; revert compression at 18; recover details at 21; sharpen at 31; reduce noise at 23; dehalo at 28; anti-alias/deblur at 42; and focus fix Off. Changed resolution to 1920x1080â âF:/CPU/01.mkvâ
Thank you very much. Thatâs pretty interesting. I still wonder if thatâs unique to the X3D variant, but Iâm not going to change my script to test it fully or enable the iGPU in bios. Someone whoâs really into it will have to take that up.
I also failed to mention that I just installed my 9900X onto an existing Windows install that used to have an Intel i5 2600K and has had like two other motherboards and processors in it. It might make a difference since I know that some reviews on the 9900X mention needing a fresh Windows install to properly turn on or off CCDsâor something like that. It sounded detrimental to me, so I didnât bother.
[Edit: I found the review and put the link here.]
Iâll try it to see if it makes a difference in speed. 
