DLA-E: a deep learning accelerator for endoscopic images classification

Journal of Big Data

Table 4 DLA-E Simulation results with MAESTRO on various DNN models and Dataflows

DNN	MobileNet V2			ResNet			VGG19
Dataflow	kcp_ws	xp_ws	rs	kcp_ws	xp_ws	rs	kcp_ws	xp_ws	rs
Number of MACs	2.98 × 10⁸	2.98 × 10⁸	2.98 × 10⁸	1.01 × 10¹⁰	1.01 × 10¹⁰	1.01 × 10¹⁰	1.77 × 10¹⁰	1.77 × 10¹⁰	1.77 × 10¹⁰
Avg L1 Size Requirement	7	7	22	8	8	37	15	15	59
Max L1 Size Requirement	18	18	96	98	98	224	18	18	96
Avg L2 Size Requirement	977	887	1,087	765	79	289	1,577	367	705
Max L2 Size Requirement	4,608	4,608	5,408	4,608	10,766	3,528	4,608	3,996	2,720
Avg Number of Utilized PEs	208	76	77	254	19	23	243	54	99
Max Number of Utilized PEs	256	240	240	256	110	191	256	222	222
Avg NoC Bandwidth	49.23	44.91	43.16	30.01	13.66	13.91	40.48	6.66	10.91
Max NoC Bandwidth	158.40	143.97	143.97	159	56	52.71	159.92	25	24.16
Avg Throughput (MACs / Cycle)	33.92	22.38	20.38	41.75	18.32	28.71	39.21	30.27	96.97
Max Throughput (MACs / Cycle)	48.00	56.00	146.67	42.67	56.00	197.73	42.67	54	214.47
Total Throughput (MACs / Cycle)	3,596.52	2,373	2,418	13,026	5,717	8,959	1,490	1,150	3,685
Total Energy Consumption (X MAC Energy)	4.56 × 10⁹	1.31 × 10¹⁰	8.40 × 10⁹	1.21 × 10¹¹	3.18 × 10¹¹	2.10 × 10¹¹	1.75 × 10¹¹	2.26 × 10¹¹	1.93 × 10¹¹
Total Runtime (Cycles)	1.73 × 10⁷	4.48 × 10⁷	4.75 × 10⁷	5.00 × 10⁸	1.41 × 10⁹	1.01 × 10⁹	8.85 × 10⁸	1.30 × 10⁹	6.12 × 10⁸