Jellyfin Transcode Benchmarks

  • Updated: AMD RDNA3 iGPU system using the Z1 Extreme (7840U) and Apple Silicon system with M1 Pro (1st gen, one encoder). Apple M4 Mac mini.
  • Primarily using 4K 120Mbps (Unlimited) and “720p” (actually 1080p) 8Mbps presets
  • Use HDR tone mapping by default
  • Using Defualt CRF of 23 for AVC and 28 for HEVC.
  • I prefer subtitle burn-in since it eleminates any subtitle sync issue in browsers. For HDR contents I prefer direct streaming using 3rd party apps to deal with subtitles.

There’s a myth that encoder and decoder alone determines transcoding efficiency. However, with morden 4K HDR videos and styled/image subtitles, transcoding is in many cases a synthetic workload for CPU, GPU and media engine, especailly for low powered NAS CPUs. Here I have multiple test footages on my NAS transcoded via Jellyfin installed on these machines:

  1. Low powered machine with a 4 E-core Intel N100 CPU, 8GB DDR4 memory and the powerful Intel Gen12 media encoders.
  2. High powered machine with a 16-Core AMD Ryzen 9 7950X3D CPU, 64GB DDR5 memory and NVIDIA RTX 4090 graphics card, eliminating any possible bottleneck on CPU and GPU 3D engine.
  3. ROG Ally with AMD Ryzen Z1 Extreme, 16GB LPDDR5 memory and Radeon 780M RDNA3 GPU to simulate 7840-like AMD mini PCs. I only did a quick test of JW4 and SW2 4Kto4K subtitle burn-in on this setup.
  4. Apple MacBook Pro 2021 with M1 Pro, 32GB LPDDR5 memory and 14-core Apple GPU (single encoder) to simulate M1 Mac minis.
  5. Apple Mac mini with M4, 16GB LPDDR5 memory and 10-core Apple GPU.

Takeaway

The result is somewhat unexpected, the intel N100 is constantly having its fast media engines underloaded when burning subtitles in Linux, and virtually everywhere in Windows. In Windows ffmpeg process can easily saturate the CPU just pulling files via SMB. On the other hand, the Intel media engine, even fully loaded, is 3-4x slower than the latest NVENC, and that’s only one of the total of 2.

While Intel’s low power iGPU is still my go for always-on home media servers, a beefier setup is certainly useful if you want to serve multiple users simultanously.

The AMD setup is somewhere in between. It has a powerful 8-core Zen 4 CPU and 12CU RDNA3 GPU to eliminate any other bottlenecks. It is nearly 2x as fast as the N100 setup, though the media codec of the latter is not fully utilized due to GPU bottleneck.

Apple Silicon is a black box hardware that’s most fascinating. It’s blazing fast and wildly efficient. On same transcoding tasks, the M4 Mac mini yields over twice the framerate of the N100 system while consuming less than half the power. However, a persistent myth is somewhat validated: Apple’s Media Engine tends to output crap footage. At a 8Mbps preset, the image quality barely outperforms Intel’s 6Mbps. Apple’s Media Engine can output 1440p from 8Mbps and beyond, while other systems typically only handle 1080p or 4K. This may give it an edge with higher presets, like 15Mbps, maybe not because it’s low quality is beyond redemption. Another thing do need mention is, it’s significantly faster at encoding HEVC than H.264.

For N100

  • VAAPI makes no difference
  • Low-power encoder is slightly faster
  • VPP tone-mapping is useless unless you want some power saving
  • Tone mapping and subtitle burning may cause GPU 3D engine bottleneck and impact performance badly.
  • Audio transcoding will also saturate CPU, capped at 5-6x.
  • Encoding in HEVC has 10%-20% performance hit compared to H.264, however, it’s only a pain point if your browser supports HEVC and you want to watch 4K HDR tone-mapped content in your home network.
  • For some videos FFMPEG have to probe through the entire video if one subtitle stream has no offset (common for ASS or SRT streams), better to turn on burn-in for small videos only.
  • Not able to handle 4K in Windows.

For Nvidia Cards

  • RTX 4090 could pull over 150W power when transcoding and hit maximium frequency.
  • Since there is no 3D and CPU bottleneck, one NVENC should constantly hit full load. For certain GPUs with multiple NVENC, since it’s a single session and NVENC split frame encoding is for HEVC and AV1 only, one should not see full utilization.

For AMD 7840U-like SoCs:

  • No bottleneck on CPU or GPU
  • Draws ~25w power with CPU at 3.2GHz (turbo boost off) and GPU at 0.8-1.6GHz.

For Apple M1 family

  • Faster than the N100 by a little bit with same level of low power consumption.

  • Not cost effective but certainlly power saving.

  • Faster when encode to HEVC.

For Apple M4

  • Up to 2x faster than N100, while consuming significantly less power. Light years ahead in terms of efficiency.
  • 1080p transcoding performance is not scaled linearly. Also the ME prefers to output 1440p footage.
  • Poor image quality makes low bitrate presets barely usable.
  • When bandwidth is not super limited and fine with macOS bundled Jellyfin app, this chip is the prime choice for JF transcoding.

Intel N100 with Intel UHD Graphics, 24EUs@750MHz, Linux

JW4

  • Bitrate: 67.4Mbps

  • Purpose: Remux 4K HDR10

  • 4Kto4K, VPP tone mapping, no sub burn-in: ~50fps

  • 4Kto4K, non-VPP tone mapping, no burn-in: 57fps

  • 4Kto4K, VPP tone mapping burin-in: 27fps

    • Lower CPU and GPU (both 3D and codec) usage
    • ~1W GPU, ~7w CPU power

  • 4Kto4K, non-VPP tone mapping burin-in: 43fps

    • ~1.8w GPU, ~10w CPU power

  • 4Kto1080p, subtitle burn-in: 82fps

GunBuster

  • Bitrate: 11.7Mbps

  • Purpose: 1080 * 1440 video

  • Original: ~200fps

  • 1080pto1080p 8Mbps: ~270fps

Mazouku

  • Bitrate: 3.7Mbps

  • Purpose: 1080p SDR WEBDL anime, with pre-burnt-in subtitles

  • Original: ~230fps

Madoka

  • Bitrate: 11.7Mbps

  • Purpose: High bitrate Bluray Anime

  • Original: 117fps

  • to 8Mbps: 114fps

SW2

  • Bitrate: 41.5Mbps

  • Purpose: High Bitrate Bluray 4K HDR10

  • 4Kto4K, sub burn-in: ~50fps

  • 4Kto4K, no subtitle: ~75fps

  • 4Kto1080p 8Mbps, sub burn-in: 94fps

Euphoria

  • Bitrate: 22.7Mbps

  • Purpose: High Bitrate HDR10 TV show

  • 4Kto4K, sub burn-in: ~50fps

  • 4Kto4K, no sub: ~70fps

  • to 1080p 8Mbps, with sub: ~110fps

LongSeason (No subtitle)

  • Bitrate: 7.1Mbps

  • Purpose: Low bitrate WEBDL 4K HDR (DV)

  • 4Kto4K: ~55fps (3D engine bottleneck, video render at ~60%)

  • to 1080p 6Mbps: ~130fps

Cyberpunk

  • Bitrate: 5.7Mbps

  • Purpose: 1080p HDR

  • Original, with sub burn-in: ~120fps

  • without sub burn-in: ~220fps

Intel N100 with Intel UHD Graphics, 24EUs@750MHz, Windows

Notes

  • Windows 11 23H2 fresh install
  • New Intel drivers break tone mapping, using driver 5085.

JW4

  • Bitrate: 67.4Mbps

  • Purpose: Remux 4K HDR10

  • 4Kto4K, VPP tone mapping, no sub burn-in: ~30fps

  • 4Kto4K, burin-in: 17fps

  • 4Kto1080p, subtitle burn-in: ~40fps

GunBuster

  • Bitrate: 11.7Mbps

  • Purpose: 1080 * 1440 video

  • Original: ~200fps

  • 1080pto1080p 8Mbps: ~210fps

Mazouku

  • Bitrate: 3.7Mbps
  • Purpose: 1080p WEBDL anime, with pre-burnt-in subtitles
  • Original: ~160fps

Madoka

  • Bitrate: 11.7Mbps

  • Purpose: High bitrate Bluray Anime

  • Original, burn-in: 89fps

  • to 6Mbps: 92fps

SW2

  • Bitrate: 41.5Mbps

  • Purpose: High Bitrate Bluray 4K HDR10

  • 4Kto4K, sub burn-in: 24fps

  • 4Kto4K, no subtitle: 36fps

  • 4Kto1080p 8Mbps, sub burn-in: 52fps

Euphoria

  • Bitrate: 22.7Mbps

  • Purpose: High Bitrate HDR10 TV show

  • 4Kto4K, sub burn-in: ~20fps

  • 4Kto4K, no sub: ~30fps

  • to 1080p 8Mbps, with sub: ~50fps

LongSeason (No subtitle)

  • Bitrate: 7.1Mbps

  • Purpose: Low bitrate WEBDL 4K HDR (DV)

  • 4Kto4K: 11fps

  • to 1080p 6Mbps: 21fps

Cyberpunk

  • Bitrate: 5.7Mbps

  • Purpose: 1080p HDR

  • Original, with sub burn-in: 53fps

  • without sub burn-in: ~100fps

NVIDIA RTX 4090, 16384CUDAs @2835MHz

JW4

  • Bitrate: 67.4Mbps

  • Purpose: Remux 4K HDR10

  • 4Kto4K, tone mapping, no sub burn-in: ~200fps

  • 4Kto1080p, subtitle burn-in: ~240fps

GunBuster

  • Bitrate: 11.7Mbps

  • Purpose: 1080 * 1440 video

  • 1080pto1080p: ~800fps

  • 1080pto1080p 8Mbps: ~900fps

Mazouku

  • Bitrate: 3.7Mbps
  • Purpose: 1080p WEBDL anime, with pre-burnt-in subtitles
  • Original: 837fps

Madoka

  • Bitrate: 11.7Mbps
  • Purpose: High bitrate Bluray Anime
  • Original, burn-in: ~650fps
  • to 6Mbps: ~700fps

SW2

  • Bitrate: 41.5Mbps

  • Purpose: High Bitrate Bluray 4K HDR10

  • 4Kto4K, sub burn-in: ~250fps

  • 4Kto4K, no subtitle: ~260fps

  • 4Kto1080p 8Mbps, sub burn-in: ~400fps

Euphoria

  • Bitrate: 22.7Mbps

  • Purpose: High Bitrate HDR10 TV show

  • 4Kto4K, sub burn-in: ~250fps

  • 4Kto4K, no sub: ~260fps

  • to 1080p 8Mbps, with sub: ~300fps

LongSeason (No subtitle)

  • Bitrate: 7.1Mbps

  • Purpose: Low bitrate WEBDL 4K HDR (DV)

  • 4Kto4K: 370fps

  • to 1080p 6Mbps: 710fps

Cyberpunk

  • Bitrate: 5.7Mbps

  • Purpose: 1080p HDR

  • Original, with sub burn-in: 860fps

  • without sub burn-in: 910fps

AMD Ryzen Z1 Extreme, 12CU RDNA3@0.8-2.7GHz

JW4

  • Bitrate: 67.4Mbps

  • Purpose: Remux 4K HDR10

  • 4Kto4K, tone mapping, sub burn-in: 76fps

SW2

  • Bitrate: 41.5Mbps
  • Purpose: High Bitrate Bluray 4K HDR10
  • 4Kto4K, sub burn-in: ~110fps

Apple M1 Pro, 14-core Apple GPU

JW4

  • Bitrate: 67.4Mbps

  • Purpose: Remux 4K HDR10

  • 4Kto4K, tone mapping, sub burn-in: 48fps (to AVC), 70fps (to HEVC)

SW2

  • Bitrate: 41.5Mbps
  • Purpose: High Bitrate Bluray 4K HDR10
  • 4Kto4K, sub burn-in: ~70fps (to AVC), ~90fps (to HEVC)

Apple M4, 10-core Apple GPU

JW4

  • Bitrate: 67.4Mbps

  • Purpose: Remux 4K HDR10

  • 4Kto4K, tone mapping, sub burn-in, CPU only, AVC: 58fps

  • 4Kto4K, tone mapping, sub burn-in, GPU, AVC: 61fps

  • 4Kto4K, tone mapping, sub burn-in, GPU, Apple Video Toolbox tone mapping, AVC: 61fps

  • 4Kto4K, tone mapping, sub burn-in, GPU, HEVC: 102fps

  • 4Kto4K, tone mapping, sub burn-in, GPU, Apple Video Toolbox tone mapping, HEVC: 102fps

  • 4Kto1080p, subtitle burn-in: ~140fps

SW2

  • Bitrate: 41.5Mbps
  • Purpose: High Bitrate Bluray 4K HDR10
  • 4Kto4K, sub burn-in, HEVC: ~130fps

Euphoria

  • Bitrate: 22.7Mbps

  • Purpose: High Bitrate HDR10 TV show

  • 4Kto4K, sub burn-in:

  • 4Kto4K, no sub:

  • to 1080p 8Mbps, with sub