Amazon Web Services#
Elastic Compute Cloud (EC2)
Launch an EC2 instance and run CUDA-X.
single-node
EC2 Cluster (with Dask)
Launch a CUDA-X cluster on EC2 with Dask.
multi-node
Elastic Kubernetes Service (EKS)
Launch a CUDA-X cluster on managed Kubernetes.
multi-node
Elastic Container Service (ECS)
Launch a CUDA-X cluster on managed container service.
multi-node
Sagemaker
Launch the RAPIDS container as a Sagemaker notebook.
single-node multi-node
NVIDIA CUDA-X libraries for data science can be deployed on Amazon Web Services (AWS) in several ways. See the list of accelerated instance types below:
Cloud |
Inst. |
Inst. |
GPU |
GPU |
xGPU |
xGPU |
|---|---|---|---|---|---|---|
AWS |
P5 |
p5.48xlarge |
8 |
H100 |
80 (GB) |
640 (GB) |
AWS |
P4 |
p4d.24xlarge |
8 |
A100 |
40 (GB) |
320 (GB) |
AWS |
P3 |
p3dn.24xlarge |
8 |
V100 |
32 (GB) |
256 (GB) |
AWS |
P3 |
p3.16xlarge |
8 |
V100 |
16 (GB) |
128 (GB) |
AWS |
P3 |
p3.8xlarge |
4 |
V100 |
16 (GB) |
64 (GB) |
AWS |
P3 |
p3.2xlarge |
1 |
V100 |
16 (GB) |
16 (GB) |
AWS |
G6 |
g6.48xlarge |
8 |
L4 |
24 (GB) |
192 (GB) |
AWS |
G6 |
g6.24xlarge |
4 |
L4 |
24 (GB) |
96 (GB) |
AWS |
G6 |
gr6.8xlarge |
1 |
L4 |
24 (GB) |
24 (GB) |
AWS |
G5 |
g5.48xlarge |
8 |
A10G |
24 (GB) |
192 (GB) |
AWS |
G5 |
g5.24xlarge |
4 |
A10G |
24 (GB) |
96 (GB) |
AWS |
G5 |
g5.16xlarge |
1 |
A10G |
24 (GB) |
24 (GB) |
AWS |
G4dn |
g4dn.metal |
8 |
T4 |
16 (GB) |
128 (GB) |
AWS |
G4dn |
g4dn.12xlarge |
4 |
T4 |
16 (GB) |
64 (GB) |
AWS |
G4dn |
g4dn.xlarge |
1 |
T4 |
16 (GB) |
16 (GB) |
Examples#
Streaming Sensitive Data Detection on EKS
Deploy a Kafka streaming pipeline on EKS that filters with cuDF and detects sensitive data with Morpheus and Triton.
Measuring Performance with the One Billion Row Challenge
Generate one billion rows with cuDF and CuPy, then aggregate them with Dask and cuDF about 60x faster than pandas.
HPO Benchmarking with CUDA-X libraries and Dask
Benchmark XGBoost and random forest HPO with Optuna on GPU versus CPU EC2 instances, up to 48x faster on GPU.
HPO with dask-ml and cuml
Tune XGBoost and a cuML random forest on airline data with dask-ml grid and random search across local GPUs.
Multi-node Multi-GPU Example on AWS using dask-cloudprovider
Train a cuML random forest on NYC taxi data on a multi-node GPU Dask cluster launched on EC2 with dask-cloudprovider.
HPO Deep Dive on SageMaker
Tune XGBoost, random forest, or KMeans on airline data across parallel SageMaker jobs, then deploy the best model.