Table 2.

Comparison of the wall-clock time of DeepVariant and DeepVariant-on-Spark with different combinations of CPUs/GPUs.

Variant caller	DeepVariant							DeepVariant-on-Spark
Machine model	CPU only				CPU+GPU			CPU only			CPU+GPU
CPUa	16	32	64	96	16	32	64	32	64	128	32	64	128
GPUb	0	0	0	0	1	2	4	0	0	0	2	4	8
Sparkc	No	No	No	No	No	No	No	Yes	Yes	Yes	Yes	Yes	Yes
AdamTransform (hr)	0	0	0	0	0	0	0	0.56	0.32	0.2	0.58	0.31	0.2
SelectBAM (hr)	0	0	0	0	0	0	0	0.5	0.33	0.23	0.48	0.29	0.2
Make_examples (hr)	6.13	3.15	1.73	1.2	5.93	3.1	1.6	2.72	1.6	1	2.82	1.48	0.83
Call_variants (hr)	10.8	6.53	5.35	3.83	1.51	1.52	1.5	3.66	2.02	0.98	0.7	0.38	0.21
Postprocess_variants (hr)	0.56	0.54	0.53	0.48	0.46	0.46	0.45	0.2	0.13	0.07	0.2	0.1	0.06
Merge VCF (hr)	0	0	0	0	0	0	0	0.02	0.02	0.02	0.02	0.02	0.02
Total time (hr)	17.49	10.22	7.61	5.51	7.9	5.08	3.55	7.66	4.42	2.5	4.8	2.58	1.52
USD/per genome	14.02	15.94	20.77	25.31	17.86	22.72	31.76	23.25	23.98	25.54	28.57	29.23	33.17
#genomes/300USDd	21	18	14	11	16	13	9	12	12	11	10	10	9

^aCPU means the number of CPU cores. ^bGPU means the number of NVIDIA Tesla P100 GPU. ^cSpark means using Apache Spark or not. ^d#genomes/300USD means the numbers of whole-genome sequence jobs that can be completed under the trial credit of 300 USD.