Performance Benchmarks

This page provides a centralized reference for evaluating AI inference performance across MediaTek Genio platforms, which aggregates benchmarking results for various AI workloads, including analytical AI and generative AI, across multiple inference frameworks such as TFLite (LiteRT) and ONNX Runtime.

Important

For more platform-specific details and comprehensive performance data, please refer to the Model Zoo.

AI Supporting Scope

The following table summarizes the AI capabilities and framework support across different MediaTek Genio platforms.

AI Supporting Scope Across Different Platform

Platform	OS	TFLite - Analytical AI (Online)	TFLite - Analytical AI (Offline)	TFLite - Generative AI	ONNX Runtime - Analytical AI
Genio 360/360P	Android	CPU + GPU + NPU	NPU	NPU	CPU + NPU
	Yocto	CPU + GPU + NPU	NPU	NPU	CPU + NPU
Genio 420/520/720	Android	CPU + GPU + NPU	NPU	NPU	CPU + NPU
	Yocto	CPU + GPU + NPU	NPU	NPU	CPU + NPU
Genio 510/700	Android	CPU + GPU + NPU	NPU	X	X
	Yocto	CPU + GPU + NPU	NPU	X	CPU
	Ubuntu	CPU + GPU + NPU	NPU	X	X
Genio 1200	Android	CPU + GPU + NPU	NPU	X	X
	Yocto	CPU + GPU + NPU	NPU	X	CPU
	Ubuntu	CPU + GPU + NPU	NPU	X	X
Genio 350	Android	CPU + GPU + NPU	X	X	X
	Yocto	CPU + GPU	X	X	CPU
	Ubuntu	CPU + GPU	X	X	X
MT8875	Android	CPU + GPU + NPU	NPU	NPU	CPU + NPU
MT8883	Android	CPU + GPU + NPU	NPU	NPU	CPU + NPU
MT8893	Android	CPU + GPU + NPU	NPU	NPU	CPU + NPU

TFLite(LiteRT) - Analytical AI

The following tables list the validated TFLite analytical models and their performance across Genio platforms. The statistics were measured using offline inference with performance mode enabled.

Important

Genio 350 does not include an NPU (MDLA/APU). Values shown for Genio 350 are measured on the GPU delegate (not NeuronSDK). For all other Genio and MT platforms, the shown numbers use NeuronSDK / Neuronrt-MDLA (offline NPU inference).

Models for Detection

Genio 350

Genio 360

Genio 420

Genio 510

Genio 520

Genio 700

Genio 720

Genio 1200

MT8875

MT8883

MT8893

Task

Model Name

Source model type

Data Type

Input Size

Inference Time (ms)

Inference Time (ms)

Inference Time (ms)

Inference Time (ms)

Inference Time (ms)

Inference Time (ms)

Inference Time (ms)

Inference Time (ms)

Inference Time (ms)

Inference Time (ms)

Inference Time (ms)

Detail

Object Detection

YOLOv5s [2]

.pt

Quant8

640x640

X

8.68

5.98

16.00

6.02

10.10

5.35

19.00

–

–

3.42

Link

Object Detection

YOLOv5s [2]

.pt

Float32

640x640

X

28.25

20.15

44.90

18.49

32.30

16.66

37.00

–

–

14.40

Link

Object Detection

YOLOv8s [2]

.pt

Quant8

640x640

X

12.92

9.40

24.00

9.02

17.00

8.04

28.40

–

–

5.64

Link

Object Detection

YOLOv8s [2]

.pt

Float32

640x640

X

42.41

30.77

69.30

26.83

50.30

24.41

54.60

–

–

16.34

Link

Footnotes

[2] (1,2,3,4)

AGPL-3.0 License: YOLOv5s and YOLOv8s are licensed under AGPL-3.0 by Ultralytics. Pre-converted models are not directly provided due to copyleft license restrictions. See YOLOv5s and YOLOv8s for details.

Models for Classification

					Genio 350	Genio 360	Genio 420	Genio 510	Genio 520	Genio 700	Genio 720	Genio 1200	MT8875	MT8883	MT8893
Task	Model Name	Source model type	Data Type	Input Size	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Detail
Classification	DenseNet	.pt	Quant8	224x224	212.44	5.86	4.27	7.30	4.17	5.30	3.83	6.00	–	–	2.40	Link
Classification	DenseNet	.pt	Float32	224x224	205.45	12.07	8.76	16.30	8.84	11.30	7.72	12.40	–	–	4.47	Link
Classification	EfficientNet	.pt	Quant8	224x224	52.28	2.20	1.46	4.00	1.59	3.00	1.56	3.00	–	–	1.16	Link
Classification	EfficientNet	.pt	Float32	224x224	44.24	4.85	3.39	9.30	3.30	6.00	3.15	6.10	–	–	2.21	Link
Classification	MobileNetV2	.pt	Quant8	224x224	29.38	1.28	0.80	1.00	0.90	1.00	0.97	1.00	–	–	0.78	Link
Classification	MobileNetV2	.pt	Float32	224x224	25.85	2.60	1.64	3.00	1.78	2.00	1.71	2.10	–	–	1.29	Link
Classification	MobileNetV3	.pt	Quant8	224x224	11.56	0.91	0.64	1.00	0.59	0.40	0.74	0.90	–	–	0.64	Link
Classification	MobileNetV3	.pt	Float32	224x224	10.12	1.67	1.09	2.00	1.13	1.10	1.20	2.00	–	–	0.97	Link
Classification	ResNet	.pt	Quant8	224x224	96.02	2.25	1.52	2.00	1.46	2.00	1.48	2.10	–	–	1.08	Link
Classification	ResNet	.pt	Float32	224x224	94.41	6.20	4.51	9.30	4.04	6.30	3.86	8.40	–	–	2.56	Link
Classification	SqueezeNet	.pt	Quant8	224x224	58.02	1.55	1.11	1.00	0.99	1.00	1.11	1.00	–	–	0.86	Link
Classification	SqueezeNet	.pt	Float32	224x224	56.68	3.57	2.54	4.10	2.26	3.00	2.31	3.00	–	–	1.64	Link
Classification	VGG	.pt	Quant8	224x224	434.96	17.72	12.78	X	13.32	X	11.17	24.40	–	–	6.47	Link
Classification	VGG	.pt	Float32	224x224	427.86	69.25	40.39	80.30	38.24	56.30	33.17	50.40	–	–	19.87	Link

Models for Recognition

Genio 350

Genio 360

Genio 420

Genio 510

Genio 520

Genio 700

Genio 720

Genio 1200

MT8875

MT8883

MT8893

Task

Model Name

Source model type

Data Type

Input Size

Inference Time (ms)

Inference Time (ms)

Inference Time (ms)

Inference Time (ms)

Inference Time (ms)

Inference Time (ms)

Inference Time (ms)

Inference Time (ms)

Inference Time (ms)

Inference Time (ms)

Inference Time (ms)

Detail

Recognition

VGGFace

.pt

Quant8

224x224

X

18.21

13.07

X

13.71

X

11.50

24.40

–

–

6.58

Link

Recognition

VGGFace

.pt

Float32

224x224

430.31

71.01

40.98

81.30

39.13

56.30

33.80

49.50

–

–

20.00

Link

Community-Contributed Model

Model Name	Source model type	Data Type	Input Size	Genio 360	Genio 420	Genio 520	Gneio 720	Genio 510	Genio 700	Genio 1200
MobileNetV1	.tflite	Quant8	224x224	X	0.85	0.93	0.92	1.28	1.04	1.05
YOLOv11s-quant8	.tflite	Quant8	640x640	11.39	18.85	18.62	8.52	27.04	19.04	X

TFLite(LiteRT) - Generative AI

For Generative AI workloads, the following tables provide representative performance data for reference and platform capability validation.

Note

The following symbols are used in the performance tables below:

• -- : To be released.

• Q3/E : Support planned for Q3 (estimated).

• X : Platform does not support this model.

Unless otherwise noted, all models listed below are supported on both Android and Yocto OS. Models or sections explicitly marked Android-only are not available on Yocto.

LLM Performance Comparison

Performance Comparison (Prompt Mode / Generative Mode) (Unit: tok/s)

Model	Genio 360	Genio 420	Genio 520	Genio 720	MT8875	MT8883	MT8893
Qwen3-0.6B	318.66 / 14.22	420.38 / 21.44	526.43 / 23.04	535.94 / 22.93	–	–	–
Qwen3-1.7B	152.67 / 8.69	222.83 / 13.03	258.83 / 13.72	262.73 / 13.84	235.67 / 6.20	834.02 / 25.18	1069.16 / 23.42
Qwen3-4B	73.21 / 3.53	105.18 / 6.87	126.12 / 7.30	125.58 / 7.25	–	–	–
Qwen3-8B	X	65.04 / 4.58	79.95 / 4.77	80.17 / 4.76	–	–	–
Qwen2.5-1.5B-Instruct	199.70 / 12.21	294.22 / 17.89	340.29 / 18.99	337.06 / 19.25	406.94 / 17.13	763.66 / 20.16	1621.85 / 38.57
Qwen2.5-3B-Instruct	94.62 / 6.71	132.30 / 9.98	161.19 / 10.38	163.23 / 10.64	220.32 / 9.60	502.05 / 19.32	751.06 / 20.87
Qwen2.5-7B-Instruct	X	57.79 / 4.61	69.71 / 4.86	69.47 / 4.73	83.07 / 4.13	184.73 / 6.89	471.95 / 11.74
gemma3-1B (Text Only)	359.43 / 16.58	513.69 / 24.58	598.27 / 27.23	583.01 / 26.44	680.41 / 21.02	1125.16 / 38.87	–
gemma3-4B (Text-Only)	92.76 / 3.44	149.00 / 5.69	176.90 / 5.90	176.79 / 5.93	–	–	–
llama3.2-1B-Instruct	233.97 / 15.94	329.32 / 21.58	385.77 / 24.76	400.57 / 24.92	–	1372.57 / 43.47	2093.61 / 61.14
llama3.2-3B-Instruct	X	120.17 / 9.92	153.84 / 10.26	153.56 / 10.36	–	611.76 / 19.89	1022.95 / 25.05
llama3-8b	X	49.23 / 4.39	56.47 / 4.66	55.87 / 4.64	–	128.36 / 6.53	426.13 / 11.51
MiniCPM-2B-sft-bf16-llama-format	X	168.67 / 5.82	153.14 / 6.48	194.79 / 7.69	–	–	886.72 / 22.28
Phi-3-mini-4k-instruct	74.79 / 4.45	101.94 / 6.30	126.82 / 7.26	127.56 / 7.28	–	–	–
Phi-3.5-mini-instruct	77.82 / 3.27	111.01 / 5.30	136.84 / 6.09	136.63 / 6.29	–	–	–
DeepSeek-R1-Distill-Qwen-1.5B	183.94 / 7.60	298.68 / 11.19	341.88 / 11.70	331.12 / 11.62	–	–	1057.25 / 25.68
DeepSeek-R1-Distill-Qwen-7B	X	56.47 / 4.60	67.80 / 4.84	67.59 / 4.83	–	–	448.17 / 11.69
DeepSeek-R1-Distill-Llama-8B	X	X	X	36.65 / 4.58	–	–	425.79 / 11.36

VLM Performance Comparison

Performance Comparison (ViT Inference Time (s) / Prompt (tok/s) / Generative (tok/s))

Model	Genio 360	Genio 420	Genio 520	Genio 720	MT8875	MT8883	MT8893
Qwen3VL-2B	0.66 / 120.62 / 5.17	0.49 / 173.16 / 7.42	0.42 / 124.25 / 7.88	0.43 / 199.75 / 8.09	–	–	–
InternVL3-1B	2.99 / 49.25 / 3.10	2.42 / 69.37 / 6.00	1.77 / 80.49 / 6.20	1.79 / 79.84 / 6.35	–	–	0.51 / 183.64 / 14.09

Speech Recognition Performance Comparison

Support Status

Model	Genio 360	Genio 420	Genio 520	Genio 720	MT8875	MT8883	MT8893
Whisper	Q3/E	Q3/E	Q3/E	Q3/E	Q3/E	Q3/E	Q3/E

Android-only Models

The following model categories are supported on Android only and are not available on Yocto.

LLM — Android-only Models

Performance Comparison (Prompt Mode / Generative Mode) (Unit: tok/s)

Model	Genio 360	Genio 420	Genio 520	Genio 720	MT8875	MT8883	MT8893
llava1.5-7b-speculative-decoding	X	58.62 / 2.99	73.12 / 3.40	73.11 / 3.40	–	138.76 / 4.46	267.98 / 6.78
medusa_v1_0_vicuna_7b_v1.5	X	X	X	91.82 / 10.56	–	–	501.05 / 22.79
vicuna1.5-7b-tree-speculative-decoding-plus	X	X	X	84.90 / 12.65	–	–	454.58 / 22.72

Stable Diffusion Performance Comparison

Performance Comparison (Main Time / Inference Time) (Unit: ms)

Model	Genio 360	Genio 420	Genio 520	Genio 720	MT8875	MT8883	MT8893
Stable Diffusion v2.1 base model with controlnet	216219 / 143784	39995 / 38425	35480 / 30421	35461 / 30365	–	–	–
Stable Diffusion v.1.5 controlnet	65449 / 53267	43507 / 42111	34483 / 33057	33870 / 32763	–	–	–

CLIP Performance Comparison

Performance Comparison (Main Time / Inference Time) (Unit: ms)

Model	Genio 360	Genio 420	Genio 520	Genio 720	MT8875	MT8883	MT8893
img_encoder_proj_clip_vit_large_dynamic	1362.52 / 435.94	619.67 / 421.55	488.69 / 314.31	473.82 / 309.29	–	–	358.61 / 51.14
img_encoder_proj_openclip_vit_big_g_dynamic	14997.09 / 5177.00	6371.25 / 5221.64	12595.05 / 4092.66	4870.33 / 3974.65	–	–	1390.56 / 517.13
img_encoder_proj_openclip_vit_h_dynamic	2440.64 / 1457.64	1760.07 / 1392.31	1376.49 / 1023.27	1331.19 / 1026.39	–	–	591.93 / 147.47
text_encoder_clip_vit_large	755.48 / 65.62	408.53 / 58.81	366.14 / 44.80	297.77 / 42.24	–	–	308.72 / 18.94
text_encoder_openclip_vit_h	1857.49 / 203.44	794.72 / 149.17	679.66 / 128.66	607.12 / 125.44	–	–	510.92 / 48.49

ONNX Runtime - Analytical AI

Important

MediaTek currently provides more comprehensive hardware-acceleration coverage for FP16 models in ONNX Runtime. Support for QDQ (INT8) quantization is under active development, and performance metrics will continue to optimize as operator coverage expands in future software releases.

The following tables list ONNX models validated on Genio platforms. Measurements were obtained using the NPU Execution Provider (where available) with performance mode enabled.

Note

The following symbols are used in the tables below:

• -- : To be released.

• X : Platform does not support this model.

Models for TAO Related

Task	Model Name	Data Type	Input Size	G520 NPU (ms)	G520 CPU (ms)	G720 NPU (ms)	G720 CPU (ms)	Pretrained Model
Object Detection	PeopleNet (ResNet34)	Float32	3x544x960	83.51	3460.66	77.63	3422.34	Download
Object Detection	PeopleNet (ResNet34)	Quant8	3x544x960	20.49	911.40	19.44	901.13	Download
Recognition	Action Recognition Net (ResNet18)	Float32	96x224x224	16.83	353.46	14.75	350.78	Download
Pose Estimation	BodyPoseNet	Float32	224x320x3	43.99	1845.36	40.56	1825.80	Download
Object Detection	LPDNet (USA Pruned)	Float32	3x480x640	3.68	107.06	3.42	106.47	Download
Segmentation	PeopleSemSegNet_AMR	Float32	3x576x960	X	7685.97	X	7749.61	Download
Segmentation	PeopleSemSegNet_AMR (Rel)	Float32	3x544x960	15.05	146.83	13.29	136.69	Download
Segmentation	PeopleSemSegNet (ShuffleSeg)	Float32	3x544x960	15.08	140.38	13.41	136.57	Download
Segmentation	PeopleSemSegNet (Vanilla Unet)	Float32	3x544x960	178.51	7510.24	163.45	7346.95	Download
Re-Identification	ReIdentificationNet (ResNet50)	Float32	3x256x128	8.59	237.53	6.87	234.10	Download
OCR	Ocrnet_resnet50	Float32	1x32x100	20.64	300.25	18.16	296.89	Download
OCR	Ocrnet_resnet50 (Pruned)	Float32	1x32x100	14.93	179.51	13.94	175.70	Download
OCR	ocd_resnet50	Float32	3x736x1280	169.41	6520.78	149.85	6340.07	Download
OCR	ocd_resnet50	Float32	3x640x640	76.18	2809.26	68.10	2748.75	Download
OCR	ocdnet_mixnet	Float32	3x640x640	362.87	17742.48	340.09	17436.33	Download
Classification	Pose Classification (ST-GCN)	Float32	3x300x34x1	223.89	787.40	207.00	772.19	Download
Pose Estimation	Centerpose (Chair DLA34)	Float32	3x512x512	X	3035.51	X	2946.96	Download
Pose Estimation	Centerpose (Camera FAN)	Float32	3x512x512	X	7689.45	X	7568.34	Download
Object Detection	LPDNet (CCPD Pruned)	Float32	3x1168x720	7.82	190.42	6.73	186.26	Download
Pose Estimation	Foundation Pose (Refiner)	Float32	6x160x160	64.91	682.93	60.97	674.01	Download
Pose Estimation	Foundation Pose (Score)	Float32	6x160x160	37.37	622.71	34.63	615.33	Download
Pose Estimation	Multi 3D Centerpose	Float32	3x512x512	X	3024.04	X	3005.77	Download

Models for Detection

Task	Model Name	Data Type	Input Size	G520 NPU (ms)	G520 CPU (ms)	G720 NPU (ms)	G720 CPU (ms)	Pretrained Model
Object Detection	YOLOv5s [1]	Quant8	640x640	X	225.74	X	221.29	Contact MediaTek FAE
Object Detection	YOLOv5s [1]	Float32	640x640	36.50	607.68	32.37	586.80	Contact MediaTek FAE
Object Detection	YOLOv8s [1]	Quant8	640x640	90.11	353.19	80.57	346.58	Contact MediaTek FAE
Object Detection	YOLO11s [1]	Quant8	640x640	102.15	301.50	90.99	295.32	Contact MediaTek FAE

[1] (1,2,3,4)

AGPL-3.0 License: YOLOv5s, YOLOv8s, and YOLO11s are licensed under AGPL-3.0 by Ultralytics. Pre-converted models are not directly provided due to copyleft license restrictions. For evaluation, contact your MediaTek FAE. For commercial use, obtain an Ultralytics Enterprise License or use Apache-2.0 alternatives.

Models for Classification

Task	Model Name	Data Type	Input Size	G520 NPU (ms)	G520 CPU (ms)	G720 NPU (ms)	G720 CPU (ms)	Pretrained Model
Classification	ConvNeXt	Quant8	224x224	X	516.21	X	1115.18	Download
Classification	ConvNeXt	Float32	224x224	X	1117.20	X	510.37	Download
Classification	DenseNet	Quant8	224x224	X	104.51	X	103.30	Download
Classification	DenseNet	Float32	224x224	8.46	205.29	7.49	200.32	Download
Classification	EfficientNet	Quant8	224x224	33.33	24.07	30.52	23.94	Download
Classification	EfficientNet	Float32	224x224	3.15	66.64	2.81	65.57	Download
Classification	MobileNetV2	Quant8	224x224	1.43	12.36	1.26	12.23	Download
Classification	MobileNetV2	Float32	224x224	1.75	31.69	1.47	30.41	Download
Classification	MobileNetV3	Quant8	224x224	X	6.30	X	6.16	Download
Classification	MobileNetV3	Float32	224x224	13.72	10.74	12.81	10.45	Download
Classification	ResNet	Quant8	224x224	2.04	45.87	1.78	45.08	Download
Classification	ResNet	Float32	224x224	3.81	112.00	3.49	111.24	Download
Classification	SqueezeNet	Quant8	224x224	9.36	33.08	8.38	31.96	Download
Classification	SqueezeNet	Float32	224x224	9.86	53.15	8.81	52.00	Download
Classification	VGG	Quant8	224x224	13.79	366.54	11.62	366.31	Download
Classification	VGG	Float32	224x224	37.17	902.03	32.24	889.14	Download

Models for Recognition

Task	Model Name	Data Type	Input Size	G520 NPU (ms)	G520 CPU (ms)	G720 NPU (ms)	G720 CPU (ms)	Pretrained Model
Recognition	VGGFace	Quant8	224x224	291.44	366.43	291.22	367.59	Download
Recognition	VGGFace	Float32	224x224	37.98	904.24	32.84	891.02	Download

Models for Robotic

Task	Model Name	Data Type	Input Size	G520 NPU (ms)	G520 CPU (ms)	G720 NPU (ms)	G720 CPU (ms)	Pretrained Model
Omni6DPose	scale_policy	Float32	1x3x3	X	0.18	X	0.17	Download
Diffusion Policy	model_diffusion_sampling	Float32	trajectory:1x16x12, global_cond:1x800	60.66	44.71	57.68	41.96	Download
MobileSam	mobilesam_encoder	Float32	3x448x448	X	705.14	X	694.74	Download
RegionNormalizedGrasp	anchornet	Float32	4x640x360	13.42	187.66	12.13	183.82	Download
RegionNormalizedGrasp	localnet	Float32	64x64x6	20.02	19.65	19.78	20.06	Download
YoloWorld	yoloworld_xl	Float32	3x640x640	465.61	11466.42	403.15	11214.33	Download

Performance Notes

Performance can vary depending on:

• The specific Genio platform and hardware configuration.

• The version of the board image and evaluation kit (EVK).

• The selected backend and model variant.

To obtain the most accurate performance numbers for your use case, you must run the application directly on the target platform.