QSV

Quick Sync Video (QSV) is the fixed-function media engine in Intel integrated graphics and Arc GPUs. FFmpeg exposes it as h264_qsv, hevc_qsv, vp9_qsv, av1_qsv, mpeg2_qsv and mjpeg_qsv. It is fast and sips power, which is why it is popular for home servers, Plex/Jellyfin transcoding, laptops and recording.

Hardware or software encoder?

QSV trades compression efficiency for speed and power like every hardware encoder: at the same bitrate it loses to x264, x265 and SVT-AV1 at slow presets. It is generally considered competitive with NVENC and ahead of AMF. Use it for transcodes where speed matters; use a software encoder when quality per bit does.

Which hardware supports what

Codec	Support
H.264	Every Quick Sync generation. 8-bit 4:2:0 only in FFmpeg (nv12).
HEVC	8-bit from Skylake (6th gen), 10-bit from Kaby Lake (7th gen). The Linux media driver only exposes 10-bit encode from Ice Lake.
VP9	Ice Lake onward, 8 and 10-bit, low-power mode only. Lunar Lake and Battlemage have no VP9 encoder.
AV1	Arc (Alchemist, Battlemage) and Meteor Lake or newer integrated graphics, 8 and 10-bit. Tiger Lake, Alder Lake and Raptor Lake only decode AV1.

Current FFmpeg builds QSV against libvpl (2.6 or newer); the older libmfx path is deprecated and has no av1_qsv. On Linux you also need intel-media-driver and a GPU runtime: vpl-gpu-rt for Tiger Lake and newer (including Arc), or the legacy Media SDK runtime for older iGPUs. On Windows the regular Intel graphics driver includes it.

Quality-oriented encode

ffmpeg -i input.mkv -c:v hevc_qsv -preset slower -global_quality 24 -c:a copy output.mkv

-global_quality without -maxrate selects ICQ (intelligent constant quality), QSV’s CRF-like mode. The range is 1–51 and lower is better, but the number is not comparable to x264/x265 CRF. For AV1:

ffmpeg -i input.mkv -c:v av1_qsv -preset slower -global_quality 30 -c:a copy output.mkv

For 10-bit HEVC or AV1 add -pix_fmt p010le. h264_qsv only takes nv12.

With neither -global_quality nor -b:v you get fixed-QP encoding (CQP at 26, or 128 for AV1) and a warning, which is rarely what you want.

H.264 can add -look_ahead 1 to -global_quality for LA_ICQ, which looks ahead for better bit distribution at the cost of latency.

Bitrate-targeted

ffmpeg -i input.mkv -c:v hevc_qsv -preset slower -b:v 6M -maxrate 9M -bufsize 12M -c:a copy output.mkv

FFmpeg picks the mode from the options: -b:v with -maxrate equal to it is CBR, otherwise VBR. Adding -global_quality to a VBR command turns it into QVBR, quality-targeted with the bitrate settings as limits. For better VBR on supported hardware add -extbrc 1 -look_ahead_depth 40.

Full-GPU pipeline

ffmpeg -hwaccel qsv -hwaccel_output_format qsv -i input.mkv -vf vpp_qsv=w=1920:h=1080 -c:v hevc_qsv -preset slower -global_quality 24 output.mkv

Frames stay in GPU memory from decode to encode. vpp_qsv scales and converts formats on the GPU; any filter that is not a QSV filter needs hwdownload first. With more than one GPU on Linux, create the device explicitly, for example -init_hw_device vaapi=va:/dev/dri/renderD128 -init_hw_device qsv=hw@va -filter_hw_device hw.

Options that matter

Option	What it does
-preset veryfast–veryslow	Speed/quality trade-off, veryslow is the best quality.
-global_quality	Quality target for ICQ (1–51). Combined with -look_ahead 1 (H.264) it selects LA_ICQ.
-b:v, -maxrate, -bufsize	Bitrate mode selection: CBR when -maxrate equals -b:v, VBR otherwise.
-extbrc 1, -look_ahead_depth N	Extended bitrate control with lookahead (up to 100 frames) for CBR and VBR. Needs support from the runtime and codec.
-bf, -b_strategy, -adaptive_b, -adaptive_i	B-frame count and adaptive B/I placement. -1 leaves it to the driver.
-low_power	Selects the fixed-function (VDENC) encoder instead of the programmable one. Default auto. Some codecs and GPUs only have this path; it has a smaller feature set, so some options are rejected.
-async_depth	Frames in flight. Lower it for latency, raise it for throughput.
-qsv_params key=value:…	Passes options straight to the VPL runtime when FFmpeg has no named option.

Pitfalls

• A listed encoder only means FFmpeg was built with it. Whether AV1, 10-bit or low-power works depends on the GPU and runtime; vpl-inspect (from libvpl) lists what the runtime supports.
• The runtime can end up in a different rate-control mode than the one you asked for. -v verbose prints a “Using the … ratecontrol method” line with the mode FFmpeg requested.
• -bf and lookahead interact with latency settings; low-delay settings (-async_depth 1 -bf 0 -low_delay_brc 1) cost compression efficiency.

Full option reference

Every private option of av1_qsv in FFmpeg 9.0.2; h264_qsv and hevc_qsv have more. -1 and auto leave the setting to the oneVPL runtime, and INT_MAX is 2147483647.

Private options

Parameter	Type	Default	Meaning	Range / named values
-async_depth	int	4	Frames processed in parallel; higher improves throughput and adds latency.	1 to INT_MAX
-preset	int	0	Target usage from veryslow (1, best) to veryfast (7). 0 leaves it to the runtime. -compression_level overrides it.	0 to 7; veryfast=7, faster=6, fast=5, medium=4, slow=3, slower=2, veryslow=1
-forced_idr	boolean	false	Make every I-frame a key frame, not only those flagged as keyframes.	0 to 1
-low_power	boolean	auto	Use the fixed-function (VDEnc) encoder. auto leaves it to the runtime.	-1 to 1
-qsv_params	dictionary	not exposed	Pass oneVPL/MediaSDK parameters directly, as key=value:key=value.	—
-b_strategy	int	-1	1 uses B-frames as references (pyramid); 0 keeps them non-reference. -1 = runtime default.	-1 to 1
-adaptive_i	int	-1	Let the encoder insert I-frames at scene changes. -1 = runtime default.	-1 to 1
-adaptive_b	int	-1	Let the encoder adapt B-frame placement. -1 = runtime default.	-1 to 1
-extbrc	int	-1	Extended bitrate control, required for -look_ahead_depth with VBR/CBR. -1 = runtime default.	-1 to 1
-low_delay_brc	boolean	auto	Strictly obey the average frame size, for low-latency streaming. auto = runtime default.	-1 to 1
-max_frame_size	int	-1	Largest encoded frame, in bytes; -1 = no limit.	-1 to INT_MAX
-max_frame_size_i	int	-1	Largest encoded I-frame, in bytes; -1 = no limit.	-1 to INT_MAX
-max_frame_size_p	int	-1	Largest encoded P-frame, in bytes; -1 = no limit.	-1 to INT_MAX
-profile	int	unknown	AV1 profile; unknown lets the runtime choose.	0 to INT_MAX; unknown=0, main=1
-tile_cols	int	0	Tile columns; 0 = runtime default.	0 to 65535
-tile_rows	int	0	Tile rows; 0 = runtime default.	0 to 65535
-look_ahead_depth	int	0	Lookahead frames for VBR/CBR when -extbrc 1 is set; 0 = off.	0 to 100

Codec-specific generic initialization overrides

The wrapper replaces these generic FFmpeg defaults for this encoder. -b is in bits/s; -g and -bf count frames. -1, and 0 for -b and -refs, leave the value to the preset, driver or rate-control logic.

Generic parameter	Default
-b	0
-g	-1
-bf	-1
-refs	0