Hardware Acceleration

Alchemist detects hardware automatically, actively probes every plausible backend/codec candidate, and selects a single active device/backend with a deterministic scoring policy. Override in Settings → Hardware.

After the first successful probe, Alchemist stores a hardware detection cache keyed by OS, architecture, FFmpeg/FFprobe versions, hardware settings, and cache schema version. If that fingerprint still matches on the next boot, the server can start with the cached backend immediately while a full probe refreshes runtime state when needed.

Detection flow (auto mode)

1. Discover plausible candidates
   • Apple VideoToolbox on macOS
   • NVIDIA NVENC when NVIDIA is present
   • Intel / AMD render nodes on Linux via /sys/class/drm/renderD*
   • AMD AMF on Windows
2. Actively probe every candidate encoder with a short FFmpeg test encode
3. Group successful probes by device path / vendor
4. Choose one active device/backend using codec coverage, backend preference, and stable vendor ordering
5. Fall back to CPU only if no GPU probe succeeds and CPU fallback is enabled

Encoder support by vendor

Vendor	AV1	HEVC	H.264	Notes
NVIDIA NVENC	RTX 40 (Ada)	Maxwell+	All	Best for speed; RTX 30 (Ampere) decodes AV1 but has no AV1 encoder
Intel QSV	Arc / Meteor Lake+	6th gen+	All	Best for power efficiency; pre-Arc / pre-Meteor-Lake iGPUs have no AV1 encoder
AMD VAAPI/AMF	RDNA 3+ on compatible driver/FFmpeg stacks	Polaris+	All	Linux VAAPI / Windows AMF; HEVC/H.264 are the validated AMD paths for 0.3.0. RDNA 2 decodes AV1 but has no AV1 encoder
Apple VideoToolbox	None (CPU only)	M1+/T2	All	No Mac chip has an AV1 hardware encoder; AV1 output uses the CPU (SVT-AV1) path. Apple Silicon can decode AV1 (M3+). Binary install recommended
CPU	All	All	All	Always available

Hardware probe

Alchemist probes each encoder at startup with a test encode using a standardized 256x256 lavfi input.

See results in Settings → Hardware → Probe Log. The UI shows:

• the selected device/backend reason
• probe counts (attempted / succeeded / failed)
• per-probe summaries for success and failure
• full FFmpeg stderr for failed probes

On Linux, explicit device paths only apply to render-node backends such as VAAPI and QSV.

Cache invalidation

The hardware cache is ignored and rebuilt when any of these change:

• operating system or architecture
• FFmpeg or FFprobe version marker
• preferred vendor, device path, CPU fallback, or CPU encoding setting
• Alchemist’s hardware detection cache version

Changing hardware settings in the UI refreshes runtime hardware state and writes a new cache entry.

Vendor-specific guides

• NVIDIA (NVENC)
• Intel (QSV / VAAPI)
• AMD (VAAPI / AMF)
• Apple (VideoToolbox)
• CPU Encoding

For Docker GPU passthrough (driver toolkit, /dev/dri, group permissions), see GPU Passthrough. If your GPU isn’t detected after setup, start with Troubleshooting.