4k transcoding Server

[JimmyDB]

Hello,
since I can't manage transcoding with my very old PC, I will have to buy a new one. (some hardware you just have to let die ) Since now several people want to stream from my server, I want to put together a hardware which can manage about 10 4k streams - in worst case with transcoding. Since I will buy this new and it should last long, I would like to have AV1 deconding possibility with it.

I thought about Ubuntu + 12gen intel CPU + GeForce 4070Ti (2NVENC Engines) + Transcoding Ramdisk

I can't find much about sizing unfortunately, do any of you have this use case or similar hardware in use? Could this work like this?
Or is there already experience how many 4k streams this graphics card can handle? (I have already found the nvidia max-stream-patch)

thanks

[TheDreadPirate]

Here's my two cents. With 10 users, its pretty easy to get that many TRANSCODES. The problem is tone mapping HDR to SDR. That is very computationally expensive and is not handled by NVENC or QSV. Its a general GPU task.

For example. My i3 12100's iGPU can do ~12 1080P transcodes. But can only do one HDR to SDR tone map.

If you are going to consistently have 10 simultaneous users, here how I would do it.

OS: Ubuntu
CPU: Any recent CPU. Match with a board that can do 128GB+ of DDR5 if you HAVE to do a RAM disk (don't think this is necessary with NVMe SSD)
GPU: Intel Arc A380 or A750
RAM: 128GB if you NEED a RAM disk, 32GB if transcodes go to SSD.
SSD: Pretty much any reputable PCIe 3 NVMe SSD. 256GB, minimum.

I also suggest you pre-transcode your 4K HDR content to 1080P SDR and keep the 4K content IN A SEPARATE LIBRARY. Because of how computationally expensive tone mapping 4K HDR is, its really not suggested you mix 4K content in with the rest of your 1080P content. It should be reserved for people with the hardware to actually play 4K HDR without transcoding and tone mapping. And keeping it in a separate library helps to enforce that separation.

[JimmyDB]

hanks for your input!

I had planned to use two Libarys to avoid having to convert the 4k content to 1080p. I read out my intention can only work if no transcoding is used.
If the 4k media is designed to work without transcoding on a FireTv stick, we all use the same FireTv, would that be an option? Or would we also have to use the same TV?

I have often read about the Intel Arc's, compared to the nvidia they are much cheaper - but with the performance really equal in value?

thanks"!

[TheDreadPirate]

If the 4k media is designed to work without transcoding on a FireTv stick, we all use the same FireTv, would that be an option? Or would we also have to use the same TV?

I'm honestly not sure how that works.  I don't have much HDR content.  More importantly I don't have any HDR displays.

I have often read about the Intel Arc's, compared to the nvidia they are much cheaper - but with the performance really equal in value?

thanks"!

Intel Arc, from a transcoding perspective, can't be beat.  Supports AV1 decoding and encoding.  Nvidia is toe-to-toe in regards to raw FPS.  But QSV is king in terms of quality.

[nyanmisaka]

Judging by its VRAM size, FP32 performance and NVENC count, it can do 10 4k->4k transcodes, including HDR2SDR tone-mapping.

AV1 decoding has been supported for a long time and AV1 encoding is available in Jellyfin 10.9 unstable.

[JimmyDB]

The hardware is taking shape, I will skip the ram disk, use fast system ssd's in raid1 instead. Regarding the graphics card, I am now unsure. I had not considered the fp32 performance before:

Rtx 4070 TI:
40tflops(fp16) 40tflops(fp32) 0.63tflops(fp64)

Arc 770
34tflops(fp16) 17tflops(fp32) 0(fp64)

There is already a huge difference, especially concerning fp32. Does anyone have an experience report from one of the cards with a similar number of users/streams?

[TheDreadPirate]

From the documentation I can find, tone mapping is fp16. Another user reported that their Arc A750 could handle 5 concurrent tone mapped streams. BTW, I think you meant A750 and not A770. The performance specs you listed correlate to the A750. The A770 has an FP16 spec of 39 TFlops.

From the encoding comparisons I've seen of Intel Arc and Nvidia NVENC in the 4000 series, Intel still has a noticeable lead in quality. Then you have to factor in AV1 ENCODING support. Save space by encoding in AV1, plus save bandwidth for clients that can direct play AV1.

[toytown]

If it helps i did some basic benchmarks with my Arc A380 back in March and was able to get 7x 2160p HDR (H265) to 2160p SDR (H264) transcodes happening at the same time https://imgur.com/gallery/jwVVujj. I hope to do some more benchmarks with the next release to see how H265/AV1 stack up.

If you load up the imgur link you do see however that after a certain amount of transcodes, performance can just drop of a cliff.
10x 2160(HDR)>1080p(SDR) gives 61fps
11x 2160(HDR)>1080p(SDR) gives 15fps
Going to guess its RAM limited on the 6GB A380.

[TheDreadPirate]

If it helps i did some basic benchmarks with my Arc A380 back in March and was able to get 7x 2160p HDR (H265) to 2160p SDR (H264) transcodes happening at the same time https://imgur.com/gallery/jwVVujj.  I hope to do some more benchmarks with the next release to see how H265/AV1 stack up. 

If you load up the imgur link you do see however that after a certain amount of transcodes, performance can just drop of a cliff.
10x 2160(HDR)>1080p(SDR) gives 61fps
11x 2160(HDR)>1080p(SDR) gives 15fps
Going to guess its RAM limited on the 6GB A380.

You know.....that got me thinking.  I can do 12 <any resolution> to 1080P (H264) transcodes, but max out at one tone map.  I wonder if I need to give my iGPU more VRAM in the BIOS.

[JimmyDB]

thanks to all
I meant the Arc 770, there are apparently different version with different tflops - also good to know
The Excel is great! If such good values are possible with an Arc A380, then I'm certainly in good hands with an A770.

Anything else I should be aware of?