DARTF cuts SAM3 latency to 158ms

DARTF makes SAM3 far more deployable, but this is a specialized NVIDIA optimization stack—not a drop-in speed boost for every hardware target.

• –Cuts Jetson latency from DART FP16’s 275 ms to 158 ms while reducing energy per frame from 13.2 J to 7.7 J
• –Preserves benchmark accuracy almost perfectly: 56.01 COCO AP for INT8 versus 56.10 for FP32
• –Uses TensorRT plugins, CUDA, CUTLASS, calibration, GPTQ, and GPU-specific engine builds, raising deployment complexity
• –The biggest practical win is edge and video inference, where 20–25 FPS on an RTX 4090 makes open-vocabulary detection genuinely usable
• –DARTF extends DART’s class-sharing insight into low-level inference optimization, showing that model architecture and deployment engineering are converging