Google’s New AI Chips Split Tasks to Save Power

Google has rolled out its newest generation of special AI chips, called Tensor Processing Units or TPUs. These chips are now split into two types: one for training AI models and another for running them. The training chips, named TPU 8t, are designed to handle the heavy work of teaching an AI by adjusting billions of tiny settings. The inference chips, called TPU 8i, focus on answering questions and making predictions once the AI is trained. Using different chips for each job lets Google’s data centers use less electricity and water. Training a model needs lots of memory and fast processing, so the 8t chips are built for that.Inference is lighter; it can run on cheaper hardware, which is why the 8i chips are smaller and use less power. This idea follows a trend in cloud computing. Amazon already offers similar chips called Inferentia, and Google’s earlier TPU v5e was aimed at smaller tasks. By matching chip power to the task, companies hope to lower operating costs and reduce their environmental impact. However, it’s unclear if Google will pass these savings on to customers. The company has highlighted the green benefits but hasn’t promised lower prices for its cloud services.