Analysis with GPU
*The interpretation of the Advanced part is only available in EN.
SAW software now supports utilizing GPU resources for running analysis.
- The tissue & cell segmentation step in the
SAW countandSAW realignpipelines - The proteome & transcriptome joint analysis in the
SAW reanalyzepipeline for Stereo-CITE FF sample
are time-consuming process.
If the GPU is enabled for calculations, the task's running time will be greatly shortened.
Here we take the NVIDIA graphics card as an example to explain how to call GPU to start SAW count run.
Driver installation
SAW pipelines are run on Linux for analysis. After your Linux server is connected to one or more NVIDIA graphics cards, ensure that the appropriate driver has been installed. If not, follow the NVIDIA CUDA Installation Guide for Linux to install CUDA.
Check GPU resources
After installing CUDA, check GPU resources in your environment as:
- Please confirm CUDA Version >= 11.7
- Please confirm that there is a device memory remaining > 3 GiB and GPU utilization < 100%
$ nvidia-smi
Tue Jun 4 05:20:12 2024
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 515.65.01 Driver Version: 515.65.01 CUDA Version: 11.7 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 Tesla T4 Off | 00000000:17:00.0 Off | 0 |
| N/A 28C P8 16W / 70W | 4MiB / 15360MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
| 1 Tesla T4 Off | 00000000:65:00.0 Off | 0 |
| N/A 34C P0 26W / 70W | 1259MiB / 15360MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
| 2 Tesla T4 Off | 00000000:CA:00.0 Off | 0 |
| N/A 36C P0 28W / 70W | 641MiB / 15360MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
| 3 Tesla T4 Off | 00000000:E3:00.0 Off | 0 |
| N/A 65C P0 66W / 70W | 14439MiB / 15360MiB | 99% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| 1 N/A N/A 83410 C .../envs/cellbin2/bin/python 1255MiB |
| 2 N/A N/A 84598 C .../envs/cellbin2/bin/python 637MiB |
| 3 N/A N/A 15814 C python 14435MiB |
+-----------------------------------------------------------------------------+
In the first table NVIDIA-SMI, you can intuitively find that there are 4 GPU graphics cards and the serial number of each one, its corresponding fan speed, card name, and performance status. The remaining detailed parameters involve GPU usage, video memory usage, etc., which will not be introduced in detail here.
The second table, Processes, displays the video memory usage, process number, and GPU occupied by each process.
Start with GPU
--gpu-id is configured to detect a GPU graphics card. Based on the GPU resource distribution in your cluster environment, choose an appropriate serial number for --gpu-id. If no one is detected, SAW count will not start the run, instead of using the CPU.
Run the analysis as:
cd /saw/runs
saw count \
--id=Demo_Mouse_Brain_with_GPU \
--sn=C04144D5\
--omics=transcriptomics \
--kit-version="Stereo-seq N FFPE V1.0" \
--sequencing-type="PE75_25+59" \
--chip-mask=/saw/datasets/chip_mask/C04144D5.barcodeToPos.h5 \
--fastqs=/saw/datasets/fastqs \
--reference=/saw/datasets/reference/mouse_transcriptome \
--image-tar=/saw/datasets/image/C04144D5_SC_20240620_153.tar.gz \
--gpu-id=0 ## Currently only accepts one GPU ID input