GOOGLE'S FIRST TENSOR PROCESSING UNIT - ARCHITECTURE

Summary from the TLDR Web Dev newsletter

Google designed the TPU v1 for fast, cost-effective inference using trained neural network models at scale. Its key feature is a focus on tensor operations, specifically matrix multiplications, which are core to neural network computations. The TPU v1 is 15-30x faster than contemporary CPUs/GPUs for inference. It has 25-29x better performance per watt than GPUs.