Cloud TPU
Train and run machine learning models faster than ever before.
Empowering businesses with Google Cloud AI
Machine learning has produced business and research breakthroughs ranging from network security to medical diagnoses. We built the Tensor Processing Unit (TPU) in order to make it possible for anyone to achieve similar breakthroughs. Cloud TPU is the custom-designed machine learning ASIC that powers Google products like Translate, Photos, Search, Assistant, and Gmail. Here’s how you can put the TPU and machine learning to work accelerating your company’s success, especially at scale.
Machine learning performance and benchmarks
To see how Cloud TPU compares to other accelerators, read the blog or view the MLPerf benchmark results.
Benefits
Built for AI on Google Cloud
Cloud TPU is designed to run cutting-edge machine learning models with AI services on Google Cloud. And its custom high-speed network offers over 100 petaflops of performance in a single pod — enough computational power to transform your business or create the next research breakthrough.
Iterate faster on your ML solutions
Training machine learning models is like compiling code: you need to update often, and you want to do so as efficiently as possible. ML models need to be trained over and over as apps are built, deployed, and refined. Cloud TPU’s robust performance and low cost make it ideal for machine learning teams looking to iterate quickly and frequently on their solutions.
Proven, state-of-the-art models
You can build your own machine learning-powered solutions for many real-world use cases. Just bring your data, download a Google-optimized reference model, and start training.
Cloud TPU offering
Cloud TPU v2
Cloud TPU v3
Cloud TPU v2 Pod
Cloud TPU v3 Pod
Cloud TPU features
Model library
Get started immediately by leveraging our growing library of optimized models for Cloud TPU. These provide optimized performance, accuracy, and quality in image classification, object detection, language modeling, speech recognition, and more.
Connect Cloud TPUs to custom machine types
You can connect to Cloud TPUs from custom AI Platform Deep Learning VM Image types, which can help you optimally balance processor speeds, memory, and high-performance storage resources for your workloads.
Fully integrated with Google Cloud Platform
At their core, Cloud TPUs and Google Cloud’s data and analytics services are fully integrated with other Google Cloud Platform offerings, like Google Kubernetes Engine (GKE). So when you run machine learning workloads on Cloud TPUs, you benefit from GCP’s industry-leading storage, networking, and data analytics technologies.
Preemptible Cloud TPU
You can save money by using preemptible Cloud TPUs for fault-tolerant machine learning workloads, such as long training runs with checkpointing or batch prediction on large datasets. Preemptible Cloud TPUs are 70% cheaper than on-demand instances, making everything from your first experiments to large-scale hyperparameter searches more affordable than ever.
Our customers
The potential of using Cloud TPU pods to accelerate our deep learning research while keeping operational costs and complexity low is a big draw. It takes us now a little over 24 hours to train models on our local GPU cluster. It will take us, depending on the size of the TPU pod, anywhere from 7 hours to 15 minutes.
Ben Mabey, Vice President, Engineering, Recursion Pharmaceuticals
Since working with Google Cloud TPUs, we’ve been extremely impressed with their speed—what could normally take days can now take hours. Cloud TPUs help us move quickly by incorporating the latest navigation-related data from our fleet of vehicles and the latest algorithmic advances from the research community.
Anantha Kancherla, Head of Software, Self-Driving Level 5, Lyft
This collaboration with Google and access to TensorFlow Research Cloud TPUs provides the ability to explore new techniques and experiments at unprecedented scale.
Athon Millane, Machine Learning and Research Lead, Max Kelsen
Google Cloud TPUs are an example of innovative, rapidly evolving technology to support deep learning, and we found that moving TensorFlow workloads to TPUs has boosted our productivity by greatly reducing both the complexity of programming new models and the time required to train them.
Alfred Spector, CTO, Two Sigma
Pricing
Cloud TPU charges for using preemptible and non-preemptible (on-demand) Cloud TPU use to train machine learning models. For detailed pricing information, please view the pricing guide.
Single Cloud TPU device pricing
The following table shows the pricing per region for using a single Cloud TPU device.
| Version | On-demand | Preemptible |
|---|---|---|
| Cloud TPU v2 | $4.50 / TPU hour | $1.35 / TPU hour |
| Cloud TPU v3 | $8.00 / TPU hour | $2.40 / TPU hour |
| Version | On-demand | Preemptible |
|---|---|---|
| Cloud TPU v2 | $4.95 / TPU hour | $1.485 / TPU hour |
| Cloud TPU v3 | $8.80 / TPU hour | $2.64 / TPU hour |
| Version | On-demand | Preemptible |
|---|---|---|
| Cloud TPU v2 | $5.22 / TPU hour | $1.566 / TPU hour |
| Cloud TPU v3 | Not available in this region | Not available in this region |
Cloud TPU Pod pricing
The following table shows the pricing for using a Cloud TPU Pod slices.
| Cloud TPU v2 Pod | Evaluation Price / hr | 1-yr Commitment Price (37% discount) | 3-yr Commitment Price (55% discount) |
|---|---|---|---|
| 32-core Pod slice | $24 USD | $132,451 USD | $283,824 USD |
| 128-core Pod slice | $96 USD | $529,805 USD | $1,135,296 USD |
| 256-core Pod slice | $192 USD | $1,059,610 USD | $2,270,592 USD |
| 512-core Pod slice | $384 USD | $2,119,219 USD | $4,541,184 USD |
| Cloud TPU v3 Pod | Evaluation Price / hr | 1-yr Commitment Price (37% discount) | 3-yr Commitment Price (55% discount) |
| 32-core Pod slice | $32 USD | $176,601 USD | $378,432 USD |
To request a Cloud TPU Pod configuration or a quote for larger Cloud TPU v3 Pod slices, please contact a sales representative.
Resources
Cloud TPU tutorials, quickstarts, and docs
Next ’19: Fast and lean data science with TPUs
Cloud TPU content from AI Hub
What’s in an image: fast, accurate image segmentation with Cloud TPUs
What makes TPUs fine-tuned for deep learning?
Train TensorFlow ML models faster and at lower cost on Cloud TPU Pods
Cloud AI products comply with the SLA policies listed here. They may offer different latency or availability guarantees from other Google Cloud services.