Minkowski Engine Benchmark

MinkUNet34 (42 conv layers)

Time to process a sparse tensor

Number of Non-zero Elements (NNZ) CPU (E5) V100 A100
161890 9.189684867858887 0.1606714725494384 0.0800826549530029
323780 22.11437630653381 0.2796809673309326 0.1274714469909668
647560 48.82245087623596 0.5429468154907227 0.2160387039184570
971340 79.74007511138916 0.8115856647491455 0.3341245651245117
1295120 107.2188804149627 1.0402274131774902 0.4301359653472900
1618900 137.5616581439972 1.3221232891082764 0.5263621807098389
1942680 167.6317486763000 1.578477144241333 0.624725341796875
2266460 196.8181667327881 1.8281619548797607 0.7229833602905273
2590240 227.5360894203186 2.0409200191497803 0.8182036876678467
3237800 290.2604696750641 2.6243700981140137 1.0436716079711914

Speedup over CPU

Number of Non-zero Elements (NNZ) V100 A100
161890 57.19549788174895 114.75250006698
323780 79.07000793646057 173.48493979283
647560 89.92123995074834 225.98937130572
971340 98.25219761124805 238.65373406972
1295120 103.0725388090386 249.26741554474
1618900 104.0460139211203 261.34411472816
1942680 106.1984009637777 268.32871577475
2266460 107.6590431211183 272.23056233783
2590240 111.4870192292528 278.09223161639
3237800 110.6019573548936 278.11475128590

MinkUNet14 (25 conv layers)

Ryzen 3700X + Titan RTX

v0.5b speed up v0.4.3
Number of Non-zero Elements (Points) Time x Time
161890 0.09865355492 3.906438882 0.3853840828
323780 0.201720953 4.181743076 0.8435451984
647560 0.3909289837 4.733338619 1.850399256
971340 0.6050679684 4.824337556 2.919052124
1295120 0.8053011894 5.007231003 4.032329082
1618900 1.009372473 5.070545621 5.118069172
1942680 1.211565018 5.053071509 6.122124672

Ryzen 3700X + Titan RTX ITX

v0.5c speed up v0.5b
Number of Non-zero Elements (Points) Time x Time
161890 0.118088483 1.301948924 0.1537451744079589
323780 0.224587202 1.420111955 0.3189389705657959
647560 0.429016828 1.448982926 0.6216380596160889
971340 0.675571441 1.431851259 0.9673178195953369
1295120 0.895347356 1.433548963 1.2835242748260498
1618900 1.114722967 1.443795053 1.609431505203247
1942680 1.339769125 1.448891065 1.9411795139312744

Fully Convolutional Geometric Features FCGF (21 conv layers)

ResUNetBN2C on Ryzen 3700X + Titan RTX

v0.5b speed up v0.4.3
Number of Non-zero Elements (Points) Time x Time
5322 0.014988899 2.631752243 0.0394470691
10644 0.016107559 4.11098283 0.0662178993
21288 0.022331714 5.615388722 0.1254012585
42576 0.035487651 6.560935463 0.2328321934
85152 0.062259197 7.109369136 0.4426236153
170304 0.113505363 7.534897789 0.8552513123
340608 0.224957943 7.712351967 1.734954834

ResUNetBN2C on Ryzen 3700X + Titan RTX ITX

v0.5c speed up v0.5b
Number of Non-zero Elements (Points) Time x Time
5322 0.0139970779 1.015772975 0.0142178535
10644 0.0171124935 1.005572968 0.0172078609
21288 0.0207984447 1.046483632 0.0217652320
42576 0.0319116115 1.075287453 0.0343141555
85152 0.0537533760 1.117707066 0.0600805282
170304 0.0964894294 1.139130628 0.1099140644
340608 0.1927807331 1.144459244 0.2206296920