Encoders

The encoder turns compositor output into a video stream. It is a per-host setting, and it decides which codecs that host can offer at all. Which codec a given session ends up using is a separate question, covered in Codecs.

What you get by default

Vulkan Video, on both AMD and NVIDIA. It is the validated path on each and there is no reason to pick anything else unless something specific pushes you off it.

On NVIDIA it also avoids an NVENC teardown fault that no driver version escapes. Vulkan is immune to it, which is why NVENC is no longer the default there.

QUASAR_ENCODER is read from the environment; nothing is auto-detected at runtime. Change it in your .env:

QUASAR_ENCODER=vulkan

Or override one machine from Admin, Fleet, Hosts, by opening that host and using its Settings page. Either way the agent restarts and live sessions on that host drop.

The four values are vulkan, va, nvenc and openh264.

Vulkan

The default. The node-agent image carries patched GStreamer Vulkan elements, so H.264, HEVC and AV1 all encode on this path.

Three switches let you turn a codec off. All default on, and you only touch them to disable something:

QUASAR_VULKAN_H264=1
QUASAR_VULKAN_HEVC=1
QUASAR_VULKAN_AV1=1

Set one to 0 and that codec moves to the vendor hardware encoder for the session instead, if the host has one. The same fallback happens by itself when the Vulkan element is missing from the image. H.264 is the exception: it is the floor and stays on Vulkan regardless, with an error in the log.

HEVC pins the compositor ring to two slots automatically, which works around a tiling defect on affected NVIDIA hardware. AV1 needs no such override.

Vulkan does not work on a driver-volume NVIDIA host

If your NVIDIA host uses automatic driver-volume provisioning rather than a full driver install, some driver versions cannot supply a Vulkan ICD the compositor can use. Set QUASAR_ENCODER=nvenc there. Admin, Fleet, Hosts reports the gap rather than claiming Vulkan works.

VA-API

Mature, and the path where zero-copy is most thoroughly worked out. Reach for it on AMD if you hit a Vulkan problem, and on Intel, where it is the better-trodden of the two options. QUASAR_ZEROCOPY=1 enables an additional full zero-copy variant that exists only here.

NVENC

NVIDIA’s own encoder. Faster to reach for historically, but the driver fault above is why it is no longer the default. Set QUASAR_ENCODER=nvenc if you want it, or if you are on a driver-volume host where Vulkan cannot work.

openh264

Software encoding, used when there is no hardware encoder. It proves an install on a machine with no GPU and that is all it is good for. Quality and throughput are both poor, and adaptive bitrate barely functions, because a software encoder cannot saturate a 1080p stream at 8 Mbps so the congestion controller never sees anything to react to.

Zero-copy

Compositor output stays on the GPU all the way to the encoder. Vulkan keeps it as Vulkan images, VA uses DMABuf, NVENC uses CUDA memory with a shared context.

This is how the pipeline is built rather than something you enable. It is worth knowing because it explains why CPU usage stays low during a session, and why the encoder is so sensitive to being pointed at the right GPU.

Picking a GPU on a multi-GPU host

QUASAR_RENDER_NODE selects the DRM render node.

ls -l /dev/dri/by-path/

Use a /dev/dri/by-path/ path rather than /dev/dri/renderD128. The numbered names are assigned in probe order and can move between reboots. By-path names follow the PCI address and do not.

For NVENC, QUASAR_CUDA_DEVICE picks the CUDA device index as well. It defaults to 0.

Tuning knobs

You should rarely need these. The defaults come from measurement.

Setting	Default	Effect
QUASAR_GOP	60	Frames between keyframes, at a 60 fps reference. Lower recovers from loss faster and costs bitrate.
QUASAR_SLICES	8	H.264 slices per frame. Quasar’s patched Vulkan encoder supports this; a stock upstream element does not.
QUASAR_TARGET_USAGE	6	VA speed against quality, 1 slowest and best, 7 fastest. Adaptive bitrate moves this under load.
QUASAR_QUEUE_BUFFERS	3	Encode queue depth. More absorbs jitter and adds latency.

Checking what you actually got

Admin, Fleet, Hosts shows the encoder each GPU reported, which is what the agent found rather than what you configured. A host set to va that reports no encoder has a problem. See Install and startup problems.

Per session, Admin, Sessions shows the encoder time series. Above roughly 16 milliseconds per frame at 60 fps, the encoder is your bottleneck.