unit	default	torch_compile_default	torch_compile_reduce_overhead	cuda_graph	cudnn	min5	A1_forced_fp32
cudal1:3090:L1:T1	1.762	1.958	2.118	1.566	1.801	1.566	1.0039212142818765
cudal1:3090:L1:T13	2.21	2.223	2.329	2.252	2.173	2.173	0.9912219953125746
cudal1:3090:L1:T17	1.857	2.032	2.09	1.927	1.775	1.775	0.9554570763176704
cudal1:a100:L1:T1	6.304	6.326	6.245	5.093	5.19	5.093	1.0129134381684581
cudal1:a100:L1:T17	5.505	5.393	5.843	5.489	5.435	5.393	0.9736666680578104
cudal1:a100:L1:T8	10.232	10.155	10.171	10.079	10.025	10.025	0.9997770351964655
cudal1:h100:L1:T1	2.75	2.457	3.777	1.896	2.9	1.896	1.0022323198992924
cudal1:h100:L1:T13	5.957	7.84	5.632	6.173	6.657	5.632	0.9948572072639214
cudal1:h100:L1:T17	2.169	3.142	3.379	1.891	2.587	1.891	0.9608602765266079
cudal1:h20:L1:T1	2.478	2.52	2.709	1.722	2.894	1.722	0.9914667363968511
cudal1:h20:L1:T13	3.352	3.625	3.774	3.458	3.544	3.352	0.9955665061436717
cudal1:h20:L1:T17	2.625	2.352	2.713	2.086	2.283	2.086	0.9746053318662229
cudal1:h20:L1:T8	4.293	4.151	3.993	3.503	4.238	3.503	0.9991805594902904
cudal1:l40:L1:T1	2.085	2.057	2.588	1.574	1.984	1.574	0.9818037495671998
cudal1:l40:L1:T13	2.356	2.204	2.691	2.632	2.399	2.204	0.9918039115780736
cudal1:l40:L1:T17	2.187	2.365	2.653	2.241	2.412	2.187	0.950311199502682
units with all available columns > 1.3x: 16/16; min over all 16x5 = 1.566; median of per-unit min = 2.180
