Table 2.

 Timing comparison between tensorQTL and LiteQTL: Times are averaged over 10 runs and expressed in seconds

	tensorQTL				LiteQTL
	Full matrix		Filtered P-value		Full matrix		Filtered max
	CPU only	CPU and GPU	CPU only	CPU and GPU	CPU only	CPU and GPU	CPU only	CPU and GPU
Data transfer	0.015	0.561	0.018	0.069	0.000	0.660	0.000	0.020
Core computation	0.940	0.055	1.601	0.029	1.022	0.054	0.536	0.030
Post processing	9.865	8.060	0.777	0.719	0.000	0.785	0.000	0.030
Elapsed	10.820	8.676	2.396	0.817	1.022	1.499	0.536	0.080

Full matrix timings are done without any filtering threshold. Filtering threshold is different for tensorQTL and LiteQTL. For tensorQTL, the MAF (Minor Allele Frequency) threshold is 0.05, and the P-value threshold is ${10}^{-5}$. For LiteQTL, the MAF threshold is 0.05, and the maximum LOD score for each transcript. The main conclusion is that the core computation and data transfer between tensorQTL and LiteQTL is very similar. The difference lies in post processing, which varies a lot depending on filtering threshold, and user-defined output.