OTT & live streaming / CDN engineering: live TV, transcoded at scale
Case study by Hasan-Uz-Zaman Ashik: GPU live transcoding for 65+ TV channels with FFmpeg and Intel Quick Sync, from SRT and multicast to an ABR ladder delivered through Broadpeak and Nimble origin packagers and CDN.
Key points
- 65+ channels on GPU: A live transcoding platform for 65+ TV channels on Intel Arc GPUs, using Quick Sync Video (QSV) for hardware decode and encode.
- Flexible ingest: Takes SRT and UDP multicast feeds, including multi-program transport streams (MPTS), and picks out each channel by program ID.
- Adaptive bitrate ladder: Every channel produces 1080p, 720p and 480p H.264 renditions for adaptive playback on any connection.
- Two delivery paths: Renditions go out over SRT to Broadpeak BkS350 and Nimble origin packagers, then through Broadpeak HPC and Nimble CDN to the players.
- Smart source handling: Interlaced sources are detected and deinterlaced automatically, sources are never upscaled, and sports and news run at 50 fps.
- One shared engine: Each channel is a small config file run by one shared script and one systemd service template, so a fix in one place applies to every channel.
- Packager-friendly encoding: A 2-second GOP with an IDR frame on every keyframe, so packagers can cut clean, aligned segments.
- Capacity planning: GPU encode and decode load is kept under about 80%, and new channels are added in small batches.
Skills & tech stack
- Streaming & video: FFmpeg, FFprobe, H.264, MPEG-TS, MPTS / SPTS, ABR ladders, Deinterlacing, GOP / IDR tuning
- Transport & ingest: SRT caller / listener, SRT latency tuning, srt-live-transmit, UDP multicast, UDP unicast
- Hardware acceleration: Intel QSV, VA-API, Intel Arc GPU, iGPU vs dGPU selection, Resizable BAR / BIOS tuning
- Delivery: Broadpeak BkS350, Broadpeak HPC, Nimble Streamer, OTT origin & packager, CDN workflows
- Networking: IP multicast, VLANs, UDP buffer & kernel tuning, Packet-loss diagnosis, tcpdump
- Linux operations: systemd templates, Bash automation, journald, nvtop, htop, mpstat
How I work
- Automate everything: Channel configs are generated from a spreadsheet, and a dry-run mode checks each one before it goes live.
- Design for failure: Independent delivery paths, automatic restarts and health checks on every channel.
- Measure before scaling: GPU, CPU and network headroom are checked before any new capacity is added.
- Document for the team: Runbooks and cheat sheets so on-call engineers can add, remove and troubleshoot channels themselves.
Troubleshooting
- Crash-looping channels (source drops, bad PIDs, restart storms)
- Wrong program from an MPTS (program ID mapping and stream selection)
- Comb artifacts on interlaced video (field order and deinterlace detection)
- Blocky video or audio clicks (packet loss, UDP buffers, NIC and kernel tuning)
- SRT connection and latency issues (caller / listener conflicts, latency budgets)
- Encoder on the wrong GPU (silent fallback between iGPU and dGPU)
- Multicast not received on a VLAN (IGMP, routing and interface binding)