Google Announces TPU v8t Sunfish and TPU v8i Zebrafish
… Where v8 trails: per-chip HBM capacity 216 GB versus Rubin’s 288 GB HBM4 , sparsity NVIDIA has a 2:4 hardware path, Google does not , and ecosystem breadth CUDA, cuDNN, TensorRT-LLM, and the PyTorch-first serving stack land on NVIDIA first; native PyTorch on TPU is still preview . …