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
GBSmall-sample 2:
12.4
GBLarge-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 ( |
2hrs: 52mins |
CPU:12; RAM:117GB |
Small-sample 2 ( |
5hrs: 15mins |
CPU:12; RAM:117GB |
Large-sample 1 ( |
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 ( |
02hrs: 52mins |
0hrs: 05mins: 4sec |
Small-sample 2 ( |
05hrs: 15mins |
0hrs: 11mins: 44sec |
Large-sample 1 ( |
89hrs: 30mins |
4hrs: 08mins |
Large-sample 2 ( |
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 ( |
4hrs: 08mins |
1hrs: 33mins |