AI-Benchmark

Have some questions regarding the scores? Faced some issues? Want to discuss the results? Welcome to our new AI Benchmark Forum!

	Power Efficiency Ranking	Power Efficiency Ranking \| Performance vs. Efficiency Tradeoff \| INT8 Results \| FP16 Results

Processor	AI Accelerator	Year	Lib	Inference Mode	INT8, FPS × (FPS per Watt)	FP16, FPS × (FPS per Watt)	Perf. – Efficiency Balance

Snapdragon 8 Gen 2	Hexagon DSP / HTP Gen 2	2022	qh.qh	FAST SINGLE ANSWER	109	24.9	52.2
Snapdragon 8 Gen 2	Hexagon DSP / HTP Gen 2	2022	qh.qh	SUSTAINED SPEED	100	24.8	50.0
Dimensity 9200	APU 690	2022	mm	FAST SINGLE ANSWER	58.6	19.3	33.6
Snapdragon 8 Gen 1	Hexagon DSP / HTP	2021	qh.qh	SUSTAINED SPEED	71	15.6	33.3
Dimensity 9200	APU 690	2022	mm	SUSTAINED SPEED	41.7	16.5	26.2
Dimensity 9000	APU 590	2021	mm	FAST SINGLE ANSWER	46.6	14.1	25.6
Dimensity 9000	APU 590	2021	mm	SUSTAINED SPEED	34.4	12	20.3
Snapdragon 888	DSP (Hexagon 780) + GPU (Adreno 660)	2020	qh.qg	SUSTAINED SPEED	64.6	5.5	18.8
Google Tensor G2	Google Tensor TPU 2.0	2022	nn	SUSTAINED SPEED	23.9	9.6	15.1
Google Tensor	Google Tensor TPU	2021	nn	SUSTAINED SPEED	12.1	8	9.8
Dimensity 800	APU 3.0 (4 cores)	2020	nn	SUSTAINED SPEED	12.7	4	7.1
Dimensity 1000+	APU 3.0 (6 cores)	2019	nn	SUSTAINED SPEED	10.2	4.4	6.7
Dimensity 820	APU 3.0 (4 cores)	2020	nn	SUSTAINED SPEED	11	3.3	6.0
Snapdragon 865	DSP (Hexagon 698) + GPU (Adreno 650)	2019	hg	SUSTAINED SPEED	8.3	4.1	5.8
Kirin 9000	GPU (Mali-G78 MP24)	2020	gg	SUSTAINED SPEED	5.2	5.5	5.3
Snapdragon 855	DSP (Hexagon 690) + GPU (Adreno 640)	2018	hg	SUSTAINED SPEED	7.1	3.3	4.8
Exynos 2100	GPU (Mali-G78 MP14)	2021	gg	SUSTAINED SPEED	4.7	4.9	4.8
Kirin 990 5G	GPU (Mali-G76 MP10)	2019	gg	SUSTAINED SPEED	4.2	4.5	4.3
Exynos 9820 Octa	GPU (Mali-G76 MP12)	2018	gg	SUSTAINED SPEED	3.5	3.7	3.6
Kirin 980	GPU (Mali-G76 MP10)	2018	gg	SUSTAINED SPEED	3.2	3.5	3.3
Snapdragon 845	DSP (Hexagon 685) + GPU (Adreno 630)	2018	hg	SUSTAINED SPEED	4	2.4	3.1
Snapdragon 710	DSP (Hexagon 685) + GPU (Adreno 616)	2018	hg	SUSTAINED SPEED	3.6	1.5	2.3

xy - This device accelerates quantized [x] and floating-point [y] models using the:

n - Android NNAPI

g - TensorFlow Lite GPU Delegate (OpenCL / OpenGL / Metal based)

h - Qualcomm Hexagon NN Direct Delegate

qh - Qualcomm QNN HTP Delegate

qd - Qualcomm QNN DSP Delegate

qg - Qualcomm QNN GPU Delegate

e - Samsung ENN Delegate

m - MediaTek Neuron Delegate

i - Apple CoreML Delegate

c - TensorFlow Lite / NNAPI default CPU backend

1 - This SoC might be using unofficial / prototype hardware or drivers

2 - These are the results of an early prototype. The results of the commercial SoC might be different

ETH Zurich, Switzerland