Source: The Register (theregister.co.uk), originally published June 2019. Some product details may have changed since this article was written.
What Are Google Cloud Deep Learning Containers?
Google Cloud Platform’s Deep Learning Containers (DLC) entered public beta in mid-2019, giving data scientists and ML engineers a quick way to launch fully configured machine learning environments — either locally via Docker or remotely on GCP infrastructure. The containers come pre-loaded with popular ML frameworks so teams can skip the painful dependency management that often slows the start of a new project.
How DLC Relates to Deep Learning VMs
Deep Learning Containers are sold alongside Google’s existing Deep Learning VMs — a collection of Debian-based disk images that provide access to NVIDIA GPUs for running ML frameworks. At the time of the beta announcement, Google offered 14 Deep Learning VM images covering TensorFlow, PyTorch, R 3.5, and Intel MKL (Math Kernel Library) with CUDA. Notably, approximately half of those VM images carried an Experimental label.
At the same time, the DLC catalogue listed around 20 container images, including TensorFlow, PyTorch, and R, each available in CPU and GPU variants. Google stated at launch that it was actively working to reach parity with all Deep Learning VM image types.
Running Deep Learning Containers Locally
One of the more practical benefits of DLC is that you can run the containers locally without incurring any immediate cloud costs. Each image ships with:
- A Python 3 environment
- The selected ML libraries (e.g., TensorFlow or PyTorch)
- A Jupyter server that starts automatically on container launch
For GPU-accelerated workloads you need nvidia-docker and a CUDA 10-compatible GPU. The general workflow is straightforward:
- Install the gcloud SDK.
- Authenticate and pull the container image you need from Google’s container registry.
- Run the container with Docker, passing the appropriate port mappings to access Jupyter in your browser.
Be aware that these are not lightweight images. At beta launch, the TensorFlow image was approximately 6 GB and the PyTorch image around 8 GB, so expect a meaningful initial download on first pull.
Jupyter as the Standard Interface
Every DLC image bundles Jupyter, the open-source interactive environment widely used in data science for combining code, equations, visualisations, and narrative text in shareable notebooks. Having Jupyter running automatically on container start means engineers can open a browser and begin experimenting immediately, without any additional setup.
Scaling Up to GCP
The local-first workflow is deliberately designed as an on-ramp to GCP. Once you are ready for more compute power, you can customise a container image — for example, adding your own Python packages — push it to the GCP Container Registry, and deploy it using several options:
- Google Kubernetes Engine (GKE) — for orchestrated, large-scale container workloads.
- Cloud Run — a serverless option for running containers without managing underlying infrastructure.
- Docker Swarm — for teams already using Swarm-based orchestration.
- AI Platform Notebooks — Google’s managed Jupyter Notebook service, where the container becomes the kernel environment.
Integration with the Broader GCP Ecosystem
Deep Learning Containers integrate with several other Google Cloud services that are common in ML pipelines:
- BigQuery — for querying large analytical datasets directly from notebooks.
- Cloud Dataproc — managed Apache Hadoop and Apache Spark clusters for large-scale data processing.
- Cloud Dataflow — batch and streaming data processing built on Apache Beam.
This tight integration is intentional: the easier it is to start locally, the more likely teams are to reach for GCP-native services as their workloads grow.
How DLC Compares to Competing Offerings
Google is not alone in this market. At the time of the beta launch, two major competitors had already established managed ML environments:
Amazon Web Services — SageMaker
Amazon SageMaker offered a mature set of ML services including pre-built Jupyter notebooks and a marketplace of models through the AWS Marketplace for Machine Learning. SageMaker’s managed training and inference infrastructure gave it a head start in enterprise adoption.
Microsoft Azure — Azure Machine Learning
Microsoft had invested heavily in Azure Machine Learning Workspaces, which guided users through building Docker images, deploying models, and creating reusable ML pipelines. Azure’s tooling had been available for a comparable period and was well integrated with the broader Microsoft ecosystem.
Cost Considerations and On-Premises Alternatives
For teams with sustained, high-volume ML workloads, running containers on your own hardware may be substantially cheaper than cloud compute. The same DLC images that run on GCP can run on-premises using Docker and nvidia-docker, making it feasible to develop and test locally and only use cloud resources for burst capacity or production inference.
Why Kubernetes Matters Here
Google’s deep involvement in the development of Kubernetes — and GKE’s status as a managed Kubernetes service — gives GCP a credible story for teams running large-scale, containerised ML workloads. Organisations already comfortable with Kubernetes for general application workloads are a natural audience for Deep Learning Containers deployed on GKE.
Summary
Google Cloud’s Deep Learning Containers lowered the barrier to entry for ML experimentation by combining a local Docker-based workflow with a clear path to GCP-scale compute. Pre-configured images with Jupyter, support for major frameworks like TensorFlow and PyTorch, and integration with GCP data services made DLC a pragmatic option for data science teams evaluating managed ML platforms in 2019. Teams considering the platform today should check the current GCP documentation for up-to-date image lists, GPU support requirements, and pricing.