qualcomm

Snapdragon 8 Gen 3 NPU

Mobile NPU benchmark target — what handheld AR & on-device LLM workloads tune against.

top models on this chip

Object detection

Sorted by latency p50 · matched-pair runs on Snapdragon 8 Gen 3 NPU

Object detection

#	Model	Quant	Latency p50	Throughput	Acc.	Power
1	YOLO27-Edge	INT8	13.1 ms	77 FPS	98.0%	6.3 W

top models on this chip

Sorted by latency p50 · matched-pair runs on Snapdragon 8 Gen 3 NPU

Depth

#	Model	Quant	Latency p50	Throughput	Acc.	Power
1	MiDaS-Distilled-S	INT8	9.53 ms	105 FPS	98.6%	7.3 W
2	ZoeDepth-Mobile	INT8	11.2 ms	89 FPS	98.5%	6.9 W

top models on this chip

Sorted by latency p50 · matched-pair runs on Snapdragon 8 Gen 3 NPU

Segmentation

#	Model	Quant	Latency p50	Throughput	Acc.	Power
1	SAM-3-Distilled	INT8	33.2 ms	30 FPS	98.0%	7.0 W

top models on this chip

Sorted by latency p50 · matched-pair runs on Snapdragon 8 Gen 3 NPU

Embedded ASR

#	Model	Quant	Latency p50	Throughput	Acc.	Power
1	Distil-Whisper-Tiny	INT8	174.3 ms	37 tok/s	97.4%	7.2 W
2	Whisper-Edge-Tiny	INT8	216.0 ms	30 tok/s	97.6%	6.8 W

top models on this chip

Sorted by latency p50 · matched-pair runs on Snapdragon 8 Gen 3 NPU

Multimodal

#	Model	Quant	Latency p50	Throughput	Acc.	Power
1	MobileCLIP-S2	INT8	21.7 ms	46 FPS	97.6%	7.4 W
2	DINO-v3-Distilled	INT8	38.2 ms	26 FPS	98.4%	6.4 W
3	Florence-2-Edge	INT8	42.3 ms	24 FPS	98.3%	7.5 W
4	MobileCLIP-B	FP16	61.5 ms	16 FPS	99.6%	7.4 W
5	SmolVLM-3B	INT8	352.8 ms	2.8 FPS	97.3%	7.3 W

top models on this chip

Sorted by latency p50 · matched-pair runs on Snapdragon 8 Gen 3 NPU

Pose

#	Model	Quant	Latency p50	Throughput	Acc.	Power
1	MoveNet-Lightning	INT8	4.72 ms	212 FPS	98.8%	7.1 W
2	MediaPipe-Hands-Edge	INT8	5.09 ms	196 FPS	98.5%	6.2 W
3	RTMPose-Edge	INT8	8.65 ms	116 FPS	98.4%	6.9 W

top models on this chip

Sorted by latency p50 · matched-pair runs on Snapdragon 8 Gen 3 NPU

OCR

#	Model	Quant	Latency p50	Throughput	Acc.	Power
1	DocLayout-Edge	INT8	9.33 ms	107 FPS	98.5%	6.6 W