Power Efficiency Ranking Power Efficiency Ranking    |    Performance vs. Efficiency Tradeoff    |    INT8 Results    |    FP16 Results
Processor AI Accelerator Year Lib Inference Mode INT8,
FPS per Watt		 FP16,
FPS per Watt		 Power Efficiency
Score
Snapdragon 8 Gen 2Hexagon DSP / HTP Gen 22022qh.qhFAST SINGLE ANSWER48.911.1 23.3
Snapdragon 8 Gen 2Hexagon DSP / HTP Gen 22022qh.qhSUSTAINED SPEED44.811.6 22.8
Dimensity 9200APU 6902022mmSUSTAINED SPEED24.89.2 15.1
Dimensity 9200APU 6902022mmFAST SINGLE ANSWER26.98.2 14.9
Snapdragon 8 Gen 1Hexagon DSP / HTP2021qh.qhSUSTAINED SPEED29.27.1 14.4
Dimensity 9000APU 5902021mmSUSTAINED SPEED21.57.3 12.5
Dimensity 9000APU 5902021mmFAST SINGLE ANSWER226.3 11.8
Snapdragon 888DSP (Hexagon 780) + GPU (Adreno 660)2020qh.qgSUSTAINED SPEED502.8 11.8
Dimensity 800APU 3.0 (4 cores)2020nnSUSTAINED SPEED15.85.2 9.1
Google Tensor G2Google Tensor TPU 2.02022nnSUSTAINED SPEED12.65.1 8.0
Dimensity 820APU 3.0 (4 cores)2020nnSUSTAINED SPEED14.74.4 8.0
Dimensity 1000+APU 3.0 (6 cores)2019nnSUSTAINED SPEED12.14.7 7.5
Google TensorGoogle Tensor TPU2021nnSUSTAINED SPEED74.3 5.5
Snapdragon 865DSP (Hexagon 698) + GPU (Adreno 650)2019hgSUSTAINED SPEED6.12.6 4.0
Snapdragon 855DSP (Hexagon 690) + GPU (Adreno 640)2018hgSUSTAINED SPEED4.92.1 3.2
Exynos 2100GPU (Mali-G78 MP14)2021ggSUSTAINED SPEED2.53 2.7
Kirin 990 5GGPU (Mali-G76 MP10)2019ggSUSTAINED SPEED2.52.6 2.5
Kirin 9000GPU (Mali-G78 MP24)2020ggSUSTAINED SPEED2.42.5 2.4
Snapdragon 845DSP (Hexagon 685) + GPU (Adreno 630)2018hgSUSTAINED SPEED3.61.4 2.2
Snapdragon 710DSP (Hexagon 685) + GPU (Adreno 616)2018hgSUSTAINED SPEED3.31.3 2.1
Kirin 980GPU (Mali-G76 MP10)2018ggSUSTAINED SPEED1.82 1.9
Exynos 9820 OctaGPU (Mali-G76 MP12)2018ggSUSTAINED SPEED1.81.9 1.8

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

Copyright © 2018-2023 by A.I.

ETH Zurich, Switzerland