Why Is Rendering On If It's Headless?

When running a simulation on a server, there's no display, so it runs in headless mode. That's where a common misconception takes hold — "No display, so no rendering, so no GPU usage either." Both of those follow-on assumptions can be wrong.

"headless" Means "No Window" — Nothing More

headless means no display window; it does not mean no rendering. Plenty of consumers need render output even when nothing is shown on screen.

Rendering diagram…

• HLS streaming / video recording — Pixels must be produced, so rendering is required.
• RTX sensors / synthetic data generation (SDG) — Render output is consumed as point clouds, images, and labels, so rendering is required.
• Pure physics batch (trajectories and KPIs only) — This is the only case where rendering is genuinely unnecessary.

A 2×2 breakdown makes the picture clear:

▶ Simulation footage — the four-camera composite from a warehouse AMR run dispatched with headless: true (overview top-left, top-down top-right, chase bottom-left, aisle bottom-right). Not a single window was opened, yet four cameras rendered at once, producing four 720p streams plus a separate 4K composite. In this run's profile, HLS capture alone consumed 22.7% of every tick — the GPU was working that hard with nothing on screen. Note the top-left overview camera framing only the roof: a useless angle still pays the full render cost.

GPU Usage Is Higher Than You Expect

⚠️ NVENC loads the GPU even with -RenderOffScreen — Encoding video in headless mode keeps NVENC (the hardware encoder) consuming substantial GPU resources. This is why "it's headless, so I can just leave it running with nohup" is risky — the GPU is working even with no visible window. Unattended runs must have an automatic shutdown (time limit) as a safety guard.

💡 WebRTC is the exception — Headless is sufficient in most cases, but some native livestream paths (WebRTC) require a display context, so headless is disabled for those. What you need to produce determines whether rendering and headless are active.

In a nutshell

📌 headless = no window ≠ no render. HLS, recording, RTX sensors, and synthetic data all require rendering in headless mode, and NVENC uses GPU resources on top of that. Rendering is truly off only for pure physics batches — unattended runs must have an automatic shutdown guard.