What changed
The NXP i.MX 95 board, equipped with a 2-TOPS neural processor and the eIQ® Neutron NPU, runs YOLO26n object detection at up to 84 FPS using EdgeFirst's optimized software.
This surpasses the stock software's performance of up to 40 FPS. The EdgeFirst optimization includes a zero-copy pipeline, a new delegate that speeds up YOLO26, and a model exported from EdgeFirst Studio instead of the default INT8 export from Ultralytics.
What this means in practice
The EdgeFirst optimizations significantly improve performance and accuracy. With the stock software, the NPU idles frequently, waiting for frames to be prepared and for operations it cannot run.
EdgeFirst's optimizations remove these waits by feeding the pipeline faster, skipping copies, splitting the model appropriately, and speeding up what the split can't reach.
These changes result in 84.1 FPS at 39% CPU usage and a 7 mAP increase in accuracy on the full COCO validation set.
Limitations and availability
The optimizations are available today at start.edgefirst.ai/nxp and will be integrated into future NXP BSPs. The EdgeFirst numbers come from a public validation session on the NXP i.MX 95 EVK. The current NXP BSP already includes early EdgeFirst integration.
The optimizations include changes in the kernel driver, the Neutron delegate, and the HAL. The EdgeFirst HAL is where the speed improvements come from, with fused kernels written for the Arm cores on the board.
The Smart Quantizer in EdgeFirst Studio splits the model just before the decode, allowing each output to keep its own range and improving accuracy. The fused kernels are not specific to YOLO and are validated for use with other models as well.
