Feature 10 of 10, Genlock Media Engine.
One clock for picture and sound.
The Genlock Media Engine is built for Apple Silicon: Metal and the media engines end to end, and light enough to leave your Mac room for the rest of the show.
Two paths, one clock: the picture and the sound are both locked to the Program Bus. The picture: sources (Cameras, screens, clips and network cameras), decode (VideoToolbox on the media engine), frames (IOSurface buffers, zero-copy), composite (Metal: colour, LUT, key in one pass), render clock (25, 30, 50 or 60 fps, its own real-time thread), outputs (CamStation Camera, every stream, the recorder, the preview). The sound: mics (Up to four, any sample rate), resample (NEON polyphase, PI drift loop), channel strip (Isolation, high-pass, gate, EQ, comp, limiter), a/v delay (Mics wait for the picture, automatically), program bus (48 kHz Float32, the program clock), mic out (CamStation Microphone). Every frame is timed against the Program Bus: sample time is the master.
One clock
The sound sets the time.
Every sound of the show is mixed on the Program Bus, inside CamStation Microphone: 48 kHz, 32-bit float. Its sample count is the program’s time. The picture has a real-time thread of its own that ticks at the show’s frame rate and draws each frame for the moment it is due, measured against that same clock, so picture and sound can’t drift apart.
CamStation Camera and CamStation Microphone report the same timing to the apps that use them: one clock domain, lips and voice together in every call. Every other device, from your mics to a network camera, runs on a clock of its own and is brought onto the Program Bus with high-quality resampling, corrected continuously.
Your voice waits for your face.
A camera’s picture takes longer to arrive than its sound. So each mic is delayed to meet the picture, by a delay measured from the real video path, about 80 ms, and clips are aligned the same way. You hear yourself without it.
A take moves everything together: a crossfade dissolves the picture and every channel of sound on one equal-power curve, crossing at half level exactly at the cut point.
25 or 50 fps where the mains are 50 Hz. Lights flicker at twice the mains frequency; a camera running at 30 or 60 fps under them shows it as a slow roll or a pulse. CamStation runs your show at 25, 30, 50 or 60 fps, so you can match the room.
Hardware all the way
Each job runs on the part of the chip built for it, and the frames never leave the memory they were made in.
VideoToolbox
Decodes your clips and network cameras, and encodes your streams and recordings, on the media engines.
Metal
Draws every frame of the program on the GPU: colour, keys and layers together, in one pass.
IOSurface
Shared frame buffers that the camera, the media engines, the GPU and the virtual camera all read: no copies, no per-frame allocation.
Neural Engine
Finds you in the picture for background blur and removal, through Vision.
CamStation runs on Apple Silicon only, from the M1 up. What still differs between chips, such as AV1 decoding or a ProRes engine, is detected at runtime, never assumed, and video is never encoded in software.
Light on your Mac
Nothing to do, nothing done.
The engine draws the program only while someone needs it: an app watching the camera, a stream, the recorder, or a preview you can see. With none of those, it sleeps, and the status bar reads Idle; only the mixer keeps the Program Bus running. A stream, an encoder or a capture exists only while it is in use.
The status bar’s last item shows what the show costs every second: the engine’s share of one core, counted the way Activity Monitor counts, its GPU time, and the power drawn. Click it for the Performance window: every part of the engine with its own cost, the work on the efficiency and performance cores, the GPU’s time against the frame’s, the frames late or missed for each output, the Mac’s heat, and one line of advice. The detail is measured only while the window is open.
Rehearse at full length on the Mac you’ll use, with every output running, and read the Performance window at the end, not the start. Heat builds slowly, and a rehearsal is the cheap way to learn your Mac needs a stand or a lower frame rate.

The numbers, and how each is known
Every fact below carries how it is known. A budget is a gate: a release doesn’t ship until it is met. A number is marked measured only when it comes from a benchmark on a release build on an M2 Pro, never from a screenshot. Most of the engine’s costs are still budgets, and we say so.
budget A hard gate a build must pass to ship.design How it works.measured Measured on a Release build on an M2 Pro.
- Runs on
- Apple Silicon (M1 or later), macOS 15.2 or later.
- Lip-sync budget
- Each mic and the clip path has a delay line set from the measured pipeline latency. Within ±45 ms at the receiving app.
- Webcam latency budget
- 2 frames or less added over the raw camera (67 ms at 30 fps).
- Every sample design
- Media plays exactly as it is: nothing is dropped, skipped, repeated or stretched to hide latency. A clip ends on its own last sample, counted by the voice that plays it. The one exception is a stream whose uplink can’t keep up even at its lowest bitrate: that platform drops whole groups of video frames until it catches up, never its sound, and CamStation shows it (DROPPING).
- Program clock design
- The CamStation Program Bus: 48 kHz, 32-bit float, stereo. The mixer runs on it directly.
- Video clock design
- A dedicated real-time thread ticking at the show’s rate: 25 or 50 fps where mains power is 50 Hz, so lights don’t flicker on camera, else 30 or 60. Frames are pulled at time t, never pushed.
- Camera and mic design
- Both virtual devices report the same zero-timestamp period and host-time anchor: one clock domain, no drift between them.
- Clock drift design
- Every other device meets the program through a lock-free ring and an adaptive polyphase resampler. Drift is resampled; a sample is never dropped or repeated.
- Takes budget
- A cut or crossfade starts within one frame of the command; every audio channel moves with the picture on an equal-power curve, crossing at −3 dB exactly at the cut point.
- Streams design
- Each platform gets its own render at its own size and rate, its own hardware encode, bitrate and reconnects. One failing never stops the others.
- Decode design
- Hardware VideoToolbox: H.264, HEVC with alpha, ProRes 4444, and AV1 on the Macs that have it, detected at runtime.
- Frames design
- IOSurface-backed pixel buffers shared by camera, media engine, GPU and the camera extension. No copies, no per-frame allocation.
- Music channel measured
- 0.02 % of one core on an M2 Pro, with both decks crossfading and the ducker working.
How this compares with other live tools on a Mac: CamStation compared with OBS, vMix and Ecamm Live.
Questions
Does CamStation run on Intel Macs?
No. It is built for Apple Silicon (M1 or later) only, on macOS 15.2 or later, so it can use those chips’ GPU, media engines and Neural Engine directly, with no slower path to fall back to.
Does CamStation ever drop frames or samples?
Never from your media: every sample of a clip, a song or your voice plays in order, and clock drift is resampled. The one exception is a stream whose uplink can’t keep up even at its lowest bitrate: that platform drops whole groups of video frames until it catches up, never its sound, and CamStation shows it. A frame the Mac draws late is counted and shown in the Performance window.
Why does the setup insist on CamStation Microphone?
Its Program Bus is the clock every sound of the show is mixed against. One clock is what keeps a two-hour show in sync to the end, and the virtual camera reports the same timing, so calls get lips and voice together.
What do BUDGET, DESIGN and MEASURED mean?
BUDGET: A hard gate a build must pass to ship. DESIGN: How it works. MEASURED: Measured on a Release build on an M2 Pro. A reading in a screenshot is none of them: it is the demo show’s.
Read on
Your show, on your Mac, the way broadcast does it.
Be the first to know when the beta opens.




