Capture here. Encode on your Mac.

Keep capture and your main application on the PC. Send its frames to a Mac’s dedicated media hardware for H.264 or HEVC encoding. This works for desktop capture, cameras, recording systems, kiosks, mini PCs, SBCs and video applications that want encoding off the source machine.

The source application sends NV12 or P010 frames → Mac encoder → H.264 or HEVC packets. It receives the encoded packets for recording or broadcast and continues to handle audio and output delivery. There is no intermediate video encode on the source PC.

FFmpeg clients run on Linux, Windows and macOS. Start with the installation guide, then choose the input your capture application supplies.

Feed raw NV12 frames

For a 1080p60 NV12 capture or frame producer:

ffmpeg -f rawvideo -pixel_format nv12 -video_size 1920x1080 -framerate 60 \
  -i capture.nv12 -an \
  -c:v h264_videotoolbox_remote -vt_remote_host "${MAC_HOST}:5555" \
  -vt_remote_wire_compression none -vt_remote_inflight 32 \
  -b:v 8M -g 120 -bf 0 output.mkv

Set MAC_HOST to the Mac endpoint. Use pipe:0 instead of capture.nv12 when your producer writes raw frames to standard input. Set the format, dimensions and frame rate to the frames it supplies. The application can mux or stream the returned packets using its normal output path.

For a Linux capture device that supplies NV12:

ffmpeg -f v4l2 -input_format nv12 -video_size 1920x1080 -framerate 60 \
  -i /dev/video0 -an \
  -c:v h264_videotoolbox_remote -vt_remote_host "${MAC_HOST}:5555" \
  -vt_remote_wire_compression none -vt_remote_inflight 32 \
  -b:v 8M -g 120 -bf 0 output.mkv

Send 10-bit frames

P010 carries 10-bit 4:2:0 video in 16-bit samples. Use p010le and HEVC Main 10:

ffmpeg -f rawvideo -pixel_format p010le -video_size 3840x2160 -framerate 30 \
  -i capture.p010 -an \
  -c:v hevc_videotoolbox_remote -vt_remote_host "${MAC_HOST}:5555" \
  -vt_remote_wire_compression none -vt_remote_inflight 32 \
  -pix_fmt p010le -profile:v main10 -b:v 20M -g 60 -bf 0 output.mkv

On Thunderbolt networking or fast Ethernet, None sends raw frames directly and saves the source CPU from compressing them. LZ4 reduces frame traffic; it is the default in FFmpeg and OBS. The format and frame rate determine the raw payload:

Video NV12 P010
1080p60 1.49 Gb/s 2.99 Gb/s
4K30 2.99 Gb/s 5.97 Gb/s
4K60 5.97 Gb/s 11.94 Gb/s

These are frame payload rates before network headers. NV12 uses 1.5 bytes per pixel; P010 uses 3. Thunderbolt networking presents an IP interface: use the Mac’s address on that link as the endpoint.

At log level 2, vtremoted reports the requested and effective socket-buffer sizes. The OS can cap the request; on macOS, inspect sysctl kern.ipc.maxsockbuf. Use these values and the link’s bandwidth-delay product when tuning system buffer limits.

Filter locally, encode remotely

An existing video input can use remote decode, local filters and remote encode. The two directions can use different compression settings. In the measured 4K filter workflow, LZ4 for decoded frames returning to Linux and no compression for uploads to the encoder delivered the highest throughput:

ffmpeg -c:v h264_videotoolbox_remote \
  -vt_remote_host "${MAC_HOST}:5555" -vt_remote_wire_compression lz4 \
  -i input.mkv -map 0:v:0 -an -vf hflip -fps_mode passthrough \
  -c:v hevc_videotoolbox_remote \
  -vt_remote_host "${MAC_HOST}:5555" -vt_remote_wire_compression none \
  -vt_remote_inflight 32 -pix_fmt nv12 -b:v 20M -g 60 -bf 0 output.mkv

For jobs whose video operations run entirely on the Mac, packet transcoding sends compressed packets in both directions.

Use your existing application

Raw frame transport

The extra bandwidth makes raw 4K frame workflows practical on a two-core mini PC. The full decode → Linux hflip → HEVC encode round trip reached 49.7 fps, compared with 50.3 fps for packet-only transcoding: 98.7% of its throughput.

All figures below are medians of three FFmpeg runs. Encode uploads NV12 and receives HEVC; decode sends H.264 and receives NV12.

3840×2160 wire compression Encode, fps Decode, fps Filter round trip, fps
None 51.2 66.5 30.4
LZ4 34.2 94.2 29.4
Zstd 25.0 55.7 22.0
LZ4 return, uncompressed upload — — 49.7

Uncompressed NV12 upload used 58% less client FFmpeg CPU time per frame than LZ4. It carried 5.09 Gb/s toward the Mac. The mixed filter round trip carried 4.96 Gb/s toward the Mac and 1.47 Gb/s returning.

Other pipeline Median fps Run range, fps
1080p NV12 upload → HEVC 188.1 188.0–188.2
1080p NV12 raw return 282.6 270.2–287.5
1080p NV12 filter round trip 147.6 138.1–149.6
4K P010 upload → HEVC Main 10 49.9 48.2–50.0
4K P010 raw return 34.1 33.9–36.6

Paced stream checks delivered every frame: 4K60 raw NV12 return, 4K30 decode/filter/encode, and 4K30 P010 upload and return. All 48 measured runs passed frame and packet checks. Each encoded pipeline was independently decoded, and all 240 checked NV12 frames were byte-identical across the three wire modes.

Transport hardware and method

Measured 2026-11-02 UTC with release v0.10.4:

Machine Configuration
Client Intel NUC7i5BNK, Intel Core i5-7260U, 8 GB RAM
Encoder Apple M2, 16 GB RAM, macOS 15.7.5
Link 20 Gb/s Thunderbolt networking, MTU 9000; single-stream TCP 12.6 Gb/s toward the Mac and 8.7 Gb/s returning

Three measured runs followed a 180-frame warm-up for each pipeline. Repeated Big Buck Bunny 2160p60 fixtures started at 60 seconds; cached raw-frame rings supplied the encode tests. HEVC output used 20 Mb/s at 4K and 8 Mb/s at 1080p, GOP 60, zero B frames and 32 in-flight frames. A fresh isolated daemon served each run. Timings include client startup, processing and output; preparation and validation run separately.

Download the transport measurements for individual repeats, CPU time, wire bytes, output hashes and paced stream checks.