Evaluation of CPU-only vs GPU-only workflows

Testing workflows using sample data sets:

  • Tested CPU-only and GPU-only workflow using two “Small-samples” and “Large-sample”

  • Raw data size:

    • Small-sample 1: 5.2 GB

    • Small-sample 2: 12.4 GB

    • Large-sample 1: 318 GB

  • Tested GPU-only workflow using additional “Large-sample” sample

    • Large-sample 2: 627 GB

Processing time

CPU-only workflow (i.e, current industry standard)

Sample (sample size)

Processing time

Resources

Small-sample 1 (5.2GB)

2hrs: 52mins

CPU:12; RAM:117GB

Small-sample 2 (12.4GB)

5hrs: 15mins

CPU:12; RAM:117GB

Large-sample 1 (318GB)

89hrs: 30mins

CPU:64; RAM:564GB

Large-sample-restricted*

19hrs: 36mins

CPU:64; RAM:564GB

  • Large-sample-restricted:

    • Restrict the tools to analysis specific set of regions that are most likely to harbour disease relevant DNA changes.

CPU-only vs GPU-only

Sample (sample size)

CPU-only

GPU-only

Small-sample 1 (5.2GB)

02hrs: 52mins

0hrs: 05mins: 4sec

Small-sample 2 (12.4GB)

05hrs: 15mins

0hrs: 11mins: 44sec

Large-sample 1 (318GB)

89hrs: 30mins

4hrs: 08mins

Large-sample 2 (627GB)

NA

7hrs: 10mins.

Performance on different NVIDIA GPU products

  • UiO ML-Nodes provides access to different GPU platforms/products

  • We implemented the GPU-pipeline on two different GPU platforms

Sample (sample size)

GeForce-RTX 3090

A100-PCIE-40GB

Large-sample 1 (318GB)

4hrs: 08mins

1hrs: 33mins