Software Impacts Supplementary Material

Runtime comparison with other SVM implementations using three different dense data sets using a single NVIDIA A100 GPU:

library	precision	small	small	medium	medium	large	large
PLSSVM	double	0.65s	100%	3.80s	98.92%	3min 44s	95.64%
PLSSVM @ commit d7e6564	double	0.44s	100%	1.34s	100%	46.5s	100%
ThunderSVM @ commit 802e6ed	double	1.16s	100%	13.51s	99.99%	7min 57s	99.99%
gpusvm @ commit 342e946	float	1.42s	92.48%	30.33s	90.78%	4min 57s	82.73%
GPUMLiB @ commit 94f8aab	float	1.94s	100%	22.93s	73.66%	15min 41s	73.26%
				5min 54s	74.37%	1h 21min	72.49%
gtSVM @ commit 44a749e	float	1.99s	100%	44.12s + 3.76s	97.85%	11min 56s + 1min 23s	95.41%

Other SVM implementations we were not able to use in our comparison due to compile or runtime errors on our Linux machine:

• DPSVM @ commit 771221f <https://github.com/thesiddharth/dpsvm/tree/771221f>: deadlock during execution
• mutisvm <https://code.google.com/archive/p/multisvm/>: Visual Studio build files only
• cuSVM <http://patternsonascreen.net/cusvm.html>: Windows 32bit only
• svm-gpu @ commit b2d1c0c <https://github.com/murtazajafferji/svm-gpu/tree/b2d1c0c>: internal Python error
• wuSVM @ commit 43be1b7 <https://github.com/aliciasun/wusvm/tree/43be1b7>: throws an exception
• SYCL-ML @ commit 7a5d5ba <https://github.com/codeplaysoftware/SYCL-ML/tree/7a5d5ba>: invalid pointer in example application
• SVMBenchmark @ commit c55d578 <https://github.com/OrcusCZ/SVMbenchmark>: different SVMs, however, in our tests none of these were able to achieve reasonable accuracy or did not execute at all

If you know fixes for one of the above SVMs or implemented a new SVM library, we would be happy to hear from you and update our table accordingly.