Frigate NVR Updates: GPU Performance and Model Support
Frigate's development focused on improving GPU compatibility with CUDA fallback fixes and expanding object detection capabilities through NanoDet-Plus model support. The changes address reliability issues when running modern YOLO models on GPU hardware.
Duration: PT1M47S
Episode overview
This episode is a short developer briefing from Frigate NVR Updates.
It explains recent repository work in plain language.
- Show: Frigate NVR Updates
- Published: 2026-06-08T13:11:17Z
- Audio duration: PT1M47S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
Good morning, this is your Frigate NVR update for June 8th, 2026.
The main story today is GPU performance reliability. Two identical pull requests from contributor nabheet tackle a critical issue where CUDA graph capture would fail catastrophically when running certain models like YOLOv8 on GPU. The problem occurs when ONNX models contain operations that can't be fully partitioned…
The fix implements graceful fallback behavior. When CUDA graph capture fails, the system now logs a warning and automatically creates a session without the enable CUDA graph option enabled, rather than crashing entirely. This change directly impacts developers running modern object detection models on GPU hardware,…
On the detection front, pull request 23422 introduces support for NanoDet-Plus models, representing a notable upgrade path from the default SSD Lite MobileNet model. The NanoDet-Plus-m-320 variant offers comparable detection quality with lower computational overhead, while the 416 variant provides improved accuracy.…
The development activity also included routine translation updates from the Weblate localization system, touching camera settings and motion search interfaces.
What this means for…