5. Available Hardware
Nvidia Tesla V100:
- 640 Tensor Cores
- 5120 CUDA Cores
- Double-Precision 7 teraFLOPS
- Single-Precision 14 teraFLOPS
- Deep Learning 112 teraFLOPS
- Interconnect Bandwidth: 32 GB/s
- Memory: 16 GB HBM2
- Max Power Consumption: 250 W
- Data center quality
Nvidia Pascal Titan Xp:
- 3840 CUDA Cores
- Memory: 12 GB GDDR5X
- Max Power Consumption: 250 W
- Consumer gaming quality
All data and pictures are from the Nvidia web site
6. How to get access
- All resources are connected to the Max Cluster
- Should be accessed via the batch system
- Documentation:
https://nagios.mdc-berlin.net/prod/wiki/doku.php?id=public:manuals:hpc:intro-en:
usage#getting_access_to_the_gpu_compute_nodes
8. Example: stardist
Example from Deep Learning Club (Dec 3rd, 2018):
Uwe Schmidt and Martin Weigert from MPI-CBG
"Deep learning based image restoration and cell segmentation for fluorescence microscopy“
Read more about their work and methods:
https://github.com/mpicbg-csbd/stardist
Use containers as an easy solution for trying out software.
Read our (short) documentation about containers on Max Cluster:
https://nagios.mdc-berlin.net/prod/wiki/doku.php?id=public:manuals:hpc:user-guide:05-containers
9. How to use it: Containers! https://ngc.nvidia.com/catalog/containers
Nvidia provides
collection of
GPU optimzed
containers with
all necessary
software
built into them
10. Using containers on Max Cluster
Create our own Singularity container from the TensorFlow Docker container with the stardist software in it.
N.B.: Needs sudo/root, i.e. use your own computer to build to container.
$ cat stardist.singularity
Bootstrap: docker
From: nvcr.io/nvidia/tensorflow:18.11-py3
%post
apt-get -y update && apt-get -y install firefox
pip install jupyter
pip install stardist
mkdir /notebooks && chmod a+rwx /notebooks
%runscript
jupyter notebook --notebook-dir=/notebooks --ip 0.0.0.0 --allow-root
$ sudo singularity build /tmp/stardist.sif stardist.singularity # Image will run on CentOS, Ubuntu, etc.
download https://github.com/mpicbg-csbd/stardist to /home/awachs/Software/stardist/
$ singularity run --nv -B /home/awachs/Software/stardist/examples:/notebooks -B /tmp:/run /tmp/stardist.sif
14. Using containers on Max Cluster
- Showed you interactive use.
- Submitting to batch system works just as well. Please consult our documentation.
- Important command to know about:
15. Trends
DL is now big business. More specialized hardware for large scale inference will appear.
Cloud providers will always offer the latest HW.
Intel:
- Xeon v6 (“Cascade Lake ”)
- Vector Neural Network Instructions (AVX-512_VNNI) to speed up inference
Nvidia:
- T4 Tensor Core GPU for AI Inference:
- Turing Tensor Cores 320
- NVIDIA CUDA® cores 2,560
- Max Power Consumption: 70 W
- Pre-Trained Networks (the ones below are for Medicine):
- NVIDIA Transfer Learning Toolkit: https://developer.nvidia.com/transfer-learning-toolkit
- NVIDIA AI-Assisted Annotation SDK: https://developer.nvidia.com/clara/annotation
Google:
- Pre-trained variant calling network: https://github.com/google/deepvariant