Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Explore a 32-minute presentation by Rose Zhu, Product Manager at Google Cloud TPU, focusing on Tensor Processing Units (TPUs) for large-scale AI training and inference. Learn about the specialized architecture of Google's TPU chips and systems, which are purpose-built ASICs designed specifically for machine learning applications. Discover how Google addresses the exponential growth of AI models through innovations in power efficiency, optical networking with Jupiter networks and ICI, and advanced liquid cooling technologies. Understand the importance of software co-design with TPUs, enabling supercomputer-like functionality supported by frameworks such as JAX and PyTorch, along with the XLA low-level compiler for maximizing performance. See real-world applications powering Google services like Gmail and YouTube, as well as external cloud customers including Anthropic, Salesforce, Mercedes, and Kakao. Get insights into the upcoming 7th generation TPU, Ironwood, scheduled for general availability in Q4 2025, with its two configurations (TPU7 and TPU7X) designed for different data center requirements. Examine Ironwood's technical specifications, including BF16 and FP8 support, teraflops performance metrics, and high bandwidth memory capabilities. Learn optimization techniques for TPU performance, including flash attention, host DRAM offload, mixed precision training, and specialized inference stacks. Understand how Google Kubernetes Engine (GKE) manages TPU orchestration with a focus on scheduling and runtime efficiency for large-scale AI workloads.
