Google Splits TPU Gen8 into Separate Training and Inference Chips

Google is bifurcating its next-generation TPU into purpose-built training and inference variants — a hardware strategy aimed squarely at NVIDIA's dominance across both workloads.