Benchmark

Speed (images/sec)

• dataset: 8000 samples
• batch size: 10, 20, 30, 40
• Optimizer: SGD
• GPU: Maxwell TITAN X (12GiB Memory)

model	batch size 10	batch size 20	batch size 30	batch size 40
CaffeNet	755.64	1054.47	1019.24	1077.63
SqueezeNet	458.27	579.37	534.68	549.55
NIN	443.88	516.21	612.83	656.14
ResNet-18	257.40	308.30	331.57	339.09
ResNet-34	149.88	182.69	201.75	207.49
Inception-BN	147.60	183.82	193.74	203.86
ResNet-50	88.55	102.44	109.98	111.04
Inception-v3	67.11	75.90	80.67	82.34
VGG16	56.38	58.01	59.80	59.35
ResNet-101	53.42	63.28	68.14	68.35
VGG19	45.02	46.88	48.62	48.28
ResNet-152	37.88	44.89	48.48	48.62
ResNet-200	22.58	25.61	27.17	27.32
ResNeXt-50	53.30	64.40	71.07	72.96
ResNeXt-101	31.76	39.56	42.90	43.99
ResNeXt-101-64x4d	18.22	23.08	out of memory	out of memory

Memory usage (MiB)

• dataset: 8000 samples
• batch size: 10, 20, 30, 40
• Optimizer: SGD
• GPU: Maxwell TITAN X (12GiB GPU Memory)

model	batch size 10	batch size 20	batch size 30	batch size 40	Reference accuracy (imagenet1k Top-5)
CaffeNet	430	496	631	716	78.3%
SqueezeNet	608	937	1331	1672	78.8%
NIN	650	902	1062	1222	81.3%
ResNet-18	814	1163	1497	1853	88.7%
ResNet-34	1127	1619	2094	2598	91.0%
Inception-BN	1007	1569	2212	2772	90.8%
ResNet-50	1875	3080	4265	5483	92.6%
Inception-v3	2075	3509	4944	6383	93.3%
VGG16	1738	2960	4751	5977	89.8%
ResNet-101	2791	4576	6341	8158	93.3%
VGG19	1920	3242	5133	6458	89.8%
ResNet-152	3790	6296	8777	11330	93.1%
ResNet-200	2051	2769	3471	4201	unknown
ResNeXt-50	2248	3863	5468	7089	93.3%
ResNeXt-101	3350	5749	8126	10539	94.1%
ResNeXt-101-64x4d	5140	8679	out of memory	out of memory	94.3%