Foundations for TPU development
Build with your preferred framework and diagnostic tools to drive peak performance
Learn about TPU architectureTensor Processing Units (TPUs) are Application-Specific Integrated Circuits (ASICs) designed by Google to accelerate machine learning workloads. Google Cloud makes TPUs available as a scalable resource.Tutorial
Understanding XLAXLA (Accelerated Linear Algebra) is a machine learning (ML) compiler that optimizes linear algebra, providing improvements in execution speed and memory usage.Guide
Accelerate production inference on TPUs
Deploy high-throughput, low-latency workloads using vLLM and optimized TPU serving stacks
Multimodal Embedding Inference on Cloud TPULearn how we bring high-dimensional vector pooling models to TPU hardware topologies to serve next-generation embedding models in production.Blog
Optimizing Qwen 3.5-397B MoE on IronwoodThis technical report details how we systematically applied a global optimization playbook to Qwen 3.5 MoE on Ironwood TPUs.Blog
Gemma 4 is now available on vLLM TPUvLLM introduces support & community recipes for Gemma 4, with Day 0 support on Google TPUs.Guide
Inference Recipes for TPUGet recipes to run the latest open weight models on Ironwood with vLLM, including Qwen3 and GPT-OSS.Code
vLLM for LLM inferenceLearn the essential skills to design and implement efficient AI inference solutions using Google Cloud's specialized hardware and popular open-source frameworks.Course
Scale model pre-training on TPUs
Achieve higher training throughput using JAX, PyTorch, and Keras on TPUs
Elastic training with MaxTextWhat happens when one machine dies in the middle of a multi-node training run? In this article we'll explore a possible answer, called elastic training, using the JAX AI stack (MaxText and Pathway) and on Cloud TPUs.Blog
Build an LLM with JAXBuild a GPT-2 style language model with 20 million parameters from scratch using JAX, the open-source library behind Google’s Gemini, Veo, and Nano Banana models.Course
How to scale your modelTraining LLMs often feels like alchemy, but understanding and optimizing the performance of your models doesn't have to. This book aims to demystify the science of scaling language models.ebook
A developer’s guide to training with Ironwood TPUsThis technical overview explores the specific methods and tools within the JAX and MaxText ecosystems designed to refine training efficiency and reach peak performance on Ironwood hardware.Guide
Inside the optimization of FP8 training on IronwoodGoogle’s Ironwood chips are our first TPUs to support 8-bit floating point precision, helping to accelerate AI training and inference by reducing memory usage and increasing throughput.Blog
MaxTextHigh performance, highly scalable, open-source LLM library and reference implementation written in pure Python/JAX and targeting Google Cloud TPUs and GPUs for training.Tutorial
Optimize post-training on TPUs
Efficiently customize and align open models for high-performance serving and deployment
Scaling Agentic RL with TunixTunix is built to maximize TPU throughput, keep environments modular, and make multi-turn training efficiency fully transparent. Here is how it works under the hood.Blog
Running PyTorch Natively on TPUs at Google ScaleTorchTPU enables developers to migrate existing PyTorch workloads to TPU with minimal code changes.Blog
TPU workload orchestration
Simplify cluster management and scale your workloads efficiently
Profile and debug TPU workloads
Uncover bottlenecks and optimize your model’s execution
Microbenchmarks for evaluating TPU performanceUse our new microbenchmark suite to evaluate TPUs for sample workloads.Guide
Deep Kernel profiling with XProfIntroducing the Kernel Profiling suite in XProf—a low-level hardware debugging suite engineered specifically for Pallas kernel authoring and optimization on Google TPUs.Blog
Advanced TPU optimization with XProfIntroducing three advanced capabilities designed to provide instruction-level insights: Continuous Profiling Snapshots, the Utilization Viewer, and LLO Bundle Visualization.Blog
Secure and scale TPU workloads
Design secure, high-performance network architectures for your AI workloads
Evolving Google’s data center networks for the AI eraHow Google created an end-to-end, vertically integrated AI technology stack that comprises everything from chips to systems, to platforms and application and agentic ecosystems.Blog
Data encryption for AI WorkloadsLearn about additional ways to encrypt your data when using Cloud Storage along with TPUs for AI workloads.Guide
Using VPC Service ControlsGet hands-on tips, guidance on how to set up a service perimeter using VPC Service Controls in the Google Cloud console.Guide
Community contributions
Resources and code samples from our developer community
Legal domain adaptation of Gemma with TunixA Tunix-based workflow for adapting Gemma to handle Indian legal domain specifications.Guide
Fine-tuning Gemma 3 to speak Gen ZAn implementation showing how to push a fine-tuning job using a single @kinetic.run() decorator.Guide
Tuning Pallas from scratchA practical walkthrough on writing, profiling, and auto-tuning custom TPU kernels.Guide
TPU developer resources
Explore official documentation, workload recipes, and the latest technical updates for Cloud TPUs
