jetson性能对比

前言

yolo, mask rcnn和其他图像分类算法在常用jetson硬件上的表现，仅供参考哦~

工程上选模型可做参考~

1，yolo系列

network	device	activation	precision	batch	DLA	framework	time
yolov3	Xavier	leaky	fp16	1	no	TRT5.1.6	24ms
-	Xavier	leaky	fp16	1	no	TRT7.1.0	18ms
-	NX	leaky	fp16	1	no	TRT7.1.0	30ms
-	TX2	leaky	fp16	1	no	TRT5.1.6	99ms
-	Xavier	leaky	fp16	4	no	TRT5.1.6	90ms(22.5ms each)
-	Xavier	leaky	fp16	4	no	TRT7.1.0	58ms(14.5ms each)
-	TX2	leaky	fp16	32	no	TRT5.1.6	2930ms(91.5ms each)
-	Xavier	leaky	fp16	32	no	TRT7.1.0	440ms(13.75ms each)
-	NX	leaky	fp16	4	no	TRT7.1.0	104ms(26ms each)
-	Xavier	leaky	int8	1	no	TRT5.1.6	20ms
-	Xavier	leaky	int8	1	no	TRT7.1.0	12.5ms
-	NX	leaky	int8	1	no	TRT7.1.0	20ms
-	Xavier	leaky	int8	4	no	TRT5.1.6	66ms(16.5ms each)
-	Xavier	leaky	int8	4	no	TRT7.1.0	36ms(9ms each)
-	Xavier	leaky	int8	32	no	TRT7.1.0	256ms(8ms each)
-	NX	leaky	int8	4	no	TRT7.1.0	64ms(16ms each)
-	Xavier	relu	fp16	4	no	TRT5.1.6	52ms(13ms each)
-	Xavier	relu	int8	1	no	TRT5.1.6	10ms
-	NX	relu	int8	1	no	TRT7.1.0	17ms
-	Xavier	relu	int8	4	no	TRT5.1.6	30ms(7.5ms each)
-	NX	relu	int8	4	no	TRT7.1.0	58ms(14.5ms each)
-	Xavier	relu	int8	4	no	TRT5.1.6	54ms(13.5ms each)
-	1050ti	relu	int8	4	no	TRT5.1.6	45ms(11.25ms each)
yolov3-tiny	Xavier	leaky	fp16	1	no	TRT5.1.6	5ms
yolo-resnet	Xavier	leaky	fp16	1	no	TRT5.1.6	14ms
-	Xavier	leaky	fp16	4	no	TRT5.1.6	44ms(11ms each)
-	Xavier	leaky	int8	1	no	TRT5.1.6	12ms
-	Xavier	leaky	int8	4	no	TRT5.1.6	39ms(10ms each)
-	Xavier	relu	fp16	4	no	TRT5.1.6	30ms(7.5ms each)
-	Xavier	relu	fp16	4	yes	TRT5.1.6	68ms(17ms each)
-	Xavier	relu	int8	4	no	TRT5.1.6	22ms(5.5ms each)
-	1050ti	relu	int8	4	no	TRT5.1.6	24ms(6ms each)

2，mask rcnn模型

device	input shape	precision	batch	framework	pure enqueue time
1050ti	1024x1024	fp32	1	TRT7.0	364ms
-	-	int8	-	-	140ms
Xavier	-	fp16	-	TRT7.1	136ms
-	-	int8	-	-	103ms
NX	-	fp32	-	-	871ms
-	-	fp16	-	-	239ms
-	-	int8	-	-	165ms

3，常用图像分类模型

network	device	precision	batch	DLA	framework	pure enqueue time
google net	apex	int8	1	no	TRT5.1.6	1.5ms
-	-	-	4	-	-	3.5ms(avg 0.9ms)
-	-	-	8	-	-	5.5ms(avg 0.7ms)
-	-	-	32	-	-	17.5ms(avg 0.55ms)
-	-	-	128	-	-	64ms(avg 0.5ms)
-	NX	half	1	-	TRT7.1	3ms
-	-	int8	-	-	-	2ms
-	tx2	half	1	no	TRT5.1.6	5.2ms
-	-	-	32	-	-	118ms(avg 3.7ms)
-	1050ti	float32	1	-	-	3.5ms
-	-	-	4	-	-	11ms(avg 2.75ms)
-	-	-	8	-	-	16ms(avg 2ms)
-	-	-	32	-	-	61ms(avg 1.9ms)
-	-	-	128	-	-	236ms(avg 1.84ms)
-	-	int8	1	-	-	1.5ms
-	-	-	4	-	-	4.5ms(avg 1.1ms)
-	-	-	8	-	-	6ms(avg 0.75ms)
-	-	-	32	-	-	24ms(avg 0.75ms)
-	-	-	128	-	-	90ms(avg 0.7ms)
resnet50	apex	int8	1	no	TRT5.1.6	2.2ms
-	-	-	4	-	-	4.3ms(avg 1.1ms)
-	-	-	8	-	-	7.5ms(avg 0.9ms)
-	-	-	32	-	-	25ms(avg 0.8ms)
-	-	-	128	-	-	94.5ms(avg 0.74ms)
-	NX	half	1	-	TRT7.1	6ms
-	-	-	32	-	-	103ms(avg 3.2ms)
-	-	int8	1	-	-	3.8ms
-	-	-	32	-	-	64ms(avg 2ms)
-	tx2	half	1	-	TRT5.1.6	13ms
-	-	-	32	-	-	320ms(avg 10ms)
-	1050ti	float32	1	no	TRT5.1.6	8ms
-	-	-	4	-	-	23ms(avg 5.75ms)
-	-	-	8	-	-	38ms(avg 4.75ms)
-	-	-	32	-	-	133ms(avg 4ms)
-	-	-	128	-	-	510ms(avg 4ms)
-	-	int8	1	-	-	3ms
-	-	-	4	-	-	8ms(avg 2ms)
-	-	-	8	-	-	14ms(avg 1.75ms)
-	-	-	32	-	-	44ms(avg 1.4ms)
-	-	-	128	-	-	167ms(avg 1.3ms)