What the graphics block will and will not encode
Quick Sync is the fixed-function media engine inside Intel processor graphics, and what it encodes depends on the generation. Xeon E3 chips with Iris Pro graphics span three generations: Haswell (the E3-1284L v3), Broadwell and Skylake. Intel's media driver and Media SDK runtime start at Broadwell, so this guide covers those two, and Intel's media driver documentation lists them separately. If lscpu reports an E3-1284L v3, the iHD driver will not load on it.
| Codec | Broadwell | Skylake |
|---|---|---|
| H.264 decode | Yes | Yes |
| H.264 encode | Yes, full-feature driver only | Yes |
| HEVC 8-bit decode | No | Yes |
| HEVC 8-bit encode | No | Yes, full-feature driver only |
| HEVC 10-bit | No | No |
| VP9 or AV1 encode | No | No |
| Largest encoded frame | 4096 x 4096 | 4096 x 4096 |
For a channel that plays H.264 over HLS, either generation does the job.
Do not guess the generation from a product name. Read it off the machine: lscpu prints the model, and Intel's specification page for that model gives the code name. The vainfo listing below is the final word on what your driver actually exposes.
The software stack, layer by layer
FFmpeg never talks to the chip directly; a missing layer shows up as an encoder that refuses to initialise.
| Layer | Ubuntu 24.04 package | Its job |
|---|---|---|
| Kernel driver | i915, in the kernel | Presents the graphics as a render node under /dev/dri/ |
| VA-API driver | intel-media-va-driver-non-free | Intel's iHD driver, full-feature build; the plain intel-media-va-driver is the free-kernel build with fewer encode paths |
| Media runtime | libmfx1 | The Intel Media SDK runtime, which is what Broadwell and Skylake use |
| Dispatcher | libvpl2 | Pulled in by FFmpeg; finds and loads the runtime |
| FFmpeg | ffmpeg | Provides h264_qsv and hevc_qsv |
The runtime is the trap. Intel's newer oneVPL GPU runtime begins at Tiger Lake, so on these Xeons the encoding is done by the older Media SDK runtime. Intel archived the Media SDK project in May 2023, so it is frozen. Ubuntu 24.04 builds FFmpeg against the newer dispatcher, and Intel documents that dispatcher as able to load the legacy library, libmfxhw64.so.1, which libmfx1 installs.
Install the layers and read the listing
On a fresh Ubuntu 24.04 install, from an account that can use sudo:
sudo apt update && sudo apt install -y ffmpeg vainfo intel-media-va-driver-non-free libmfx1
ls -l /dev/dri
LIBVA_DRIVER_NAME=iHD vainfo
ffmpeg -hide_banner -encoders | grep qsv
The non-free driver sits in Ubuntu's multiverse component and the others in universe. If apt cannot find a package, enable it with sudo add-apt-repository multiverse and run the install line again. Then work down this list:
ls -l /dev/drishows arenderDnode. No node means the operating system cannot see the graphics, and nothing higher up will work.vainfonames the Intel iHD driver and pairsVAProfileH264MainandVAProfileH264HighwithVAEntrypointEncSlice(orVAEntrypointEncSliceLP). Lines ending inVAEntrypointVLDalone are decode only.ffmpeg -encoderslistsh264_qsvandhevc_qsv. That proves the build has them, not that the chip will run them; the test in the next section does that.- Running FFmpeg as a non-root account? Note the group that owns the render node in the
ls -loutput, add the account to it, and log in again.
A test encode, then a live relay
Prove the hardware path with a synthetic picture and no output file:
ffmpeg -hide_banner -v verbose -f lavfi -i testsrc2=size=1920x1080:rate=30 -t 30 \
-vf format=nv12 -c:v h264_qsv -preset medium \
-b:v 4M -maxrate 4M -bufsize 8M -g 60 -f null -
What to expect: the progress line shows speed= above 1x and a frame count that climbs without stalls. With -v verbose the encoder prints the rate-control method the runtime really chose. FFmpeg's documentation warns that it can differ from the one you requested, so read it rather than assume.
For live work, pull, transcode, push keeps the machine closed: FFmpeg reads your source, scales on the processor, encodes on the graphics and sends RTMP onward. Your own live video channel with us can be that target; PRO and Platinum take a 5 Mbps ingest at 1080p.
ffmpeg -hide_banner -i "SOURCE_URL" \
-vf scale=-2:720,format=nv12 \
-c:v h264_qsv -preset medium -b:v 3M -maxrate 3M -bufsize 6M -g 60 \
-c:a aac -b:a 128k -ar 48000 \
-f flv "rtmp://INGEST_HOST/APP/STREAM_KEY"
Replace SOURCE_URL and the RTMP address with your own. -g 60 puts a keyframe every two seconds for a 30 fps source; set it to twice your frame rate. Confirm on the receiving end: the picture plays from the channel's HLS link.
The stream key rides on the command line, so anyone with a shell on the machine can read it in the ps listing; keep other accounts off the box. Load it from an environment file (mode 600) in your systemd unit rather than pasting it at a prompt; that keeps it out of history files and out of the unit itself.
Settings to copy
FFmpeg chooses the QSV rate-control mode from the combination of options you pass, not from one switch. Per its encoder documentation:
| You pass | Mode | Good for |
|---|---|---|
-b:v and an equal -maxrate | CBR | Live ingest that expects a steady rate |
-b:v and a higher -maxrate | VBR | Recordings and on-demand files |
-b:v with -look_ahead 1 (H.264) | VBR with lookahead | More quality per bit, with added delay |
-global_quality from 1 to 51, 1 best | ICQ | Files whose size may vary |
| None of these | CQP | Rarely what a live job wants |
-presetruns fromveryfasttoveryslow. Start atmediumand move one step at a time while you watchspeed=.-gsets the distance between keyframes;-bfsets the number of B-frames.-async_depthraises throughput at the cost of latency: more frames in flight before FFmpeg syncs.-low_powerselects the low-power encode path. Intel's tables give Broadwell none, so leave it off unlessvainfoshowsVAEntrypointEncSliceLP.
Which of our machines fits
Quick Sync needs physical Intel graphics, so this is work for a whole machine rather than a slice of one. Our dedicated servers in New York include two video encoder builds on Xeon E3 processors with Iris Pro graphics: the standard build, and a four-bay build with more memory, four drives and a 10 Gbps port. The box is single-tenant, with full root access and your choice of Linux distribution or Windows, ready in about 30 minutes. The port is unmetered: a relay running all day brings no per-gigabyte bill.
Take the four-bay build when recording or a library shares the machine. For one camera feed in front of viewers you may need no encoder server: one of our live video channels accepts an RTMP push from OBS, vMix or Wirecast, then serves it to browsers as HLS. Our cloud servers list vCPU, memory and SSD with no graphics engine, so keep Quick Sync on the encoder builds.
Keeping it on air
- Run the relay under systemd with
Restart=on-failure;systemctl enable --nowlaunches it immediately and again at every boot. - Leave SSH as the only open inbound port. With ufw, allow SSH first (
sudo ufw allow OpenSSH), check that/etc/default/ufwholdsIPV6=yesso the IPv6 side gets the same rules, thensudo ufw enable. - After any kernel, driver or FFmpeg update, rerun
vainfoand the null test before the next broadcast. - Watch
speed=in the journal: under 1x means the relay is falling behind real time.
When it goes wrong
- No node under
/dev/dri. The graphics are invisible to the operating system and no FFmpeg flag will help. Open a support ticket with the output oflspciandls -l /dev/dri. - vainfo loads the wrong driver or cannot start. Run
LIBVA_DRIVER_NAME=iHD vainfo. If that works, set the same variable in your service's environment. - vainfo lists decode only. The free-kernel driver is in place; install
intel-media-va-driver-non-free. - h264_qsv will not open while vainfo looks right. The runtime is the usual gap: confirm
libmfx1is installed. - hevc_qsv refuses on Broadwell. Expected: that generation has no HEVC encoder. Use
h264_qsv. - Frames do not stay on the graphics.
-hwaccel qsvkeeps them there only when decoder and encoder are both QSV and no filters run; a scaling relay decodes in software.
Honest limits
- 4096 x 4096 is the largest frame these generations encode in Intel's tables.
- No HEVC 10-bit, VP9 or AV1 encode. If a destination demands AV1, this is the wrong chip.
- The Media SDK runtime is archived. New Intel media features land in the oneVPL GPU runtime, which starts at Tiger Lake.
- For an archive where file size matters more than speed, compare against a software encoder on your own footage first.