3.1.5. GPUs, TPUs, and Specialized Compute
💡 First Principle: GPUs and TPUs both accelerate parallel computation, but they're optimized for fundamentally different workloads — choosing between them is a question of what kind of math your workload is actually doing, not which one is "more powerful."
GPUs (Graphics Processing Units) are general-purpose parallel processors well-suited to a broad range of workloads: graphics rendering, general machine learning training and inference, and scientific computing. TPUs (Tensor Processing Units) are Google's custom-built chips designed specifically for the matrix and tensor operations that underlie large-scale machine learning — particularly training large neural networks — and can outperform GPUs on those specific workloads while being a poor fit for anything outside that narrow lane (like general graphics rendering).
⚠️ Exam Trap: A scenario emphasizing large-scale, Google-optimized deep learning model training is a strong hint toward TPUs specifically, while a scenario describing more general-purpose parallel workloads, or a need for broader software/framework compatibility, points toward GPUs — the exam expects you to read the type of workload described, not just "needs acceleration."
Reflection Question: If a scenario describes a team training a massive, Google Cloud-native deep learning model at scale and asks which accelerator is purpose-built for that exact workload, what's the answer, and why would the same accelerator be a poor choice for a general 3D-rendering pipeline?
