TPU7x (Ironwood)

TPU7x (Ironwood) is the generation for the largest training and inference workloads. Each chip has 2 TensorCores and 4 SparseCores, and a pod holds up to 9216 chips.

Use the specification topic to size a slice, the configurations topic to choose a slice shape, and the VM types topic to see what each VM provides.

Important: The Cloud TPU API is no longer under active development and is not supported on TPU7x (Ironwood) and later. Create TPU7x (Ironwood) capacity with Google Kubernetes Engine (GKE) or Compute Engine.

Note: JAX and PyTorch are supported on TPU7x (Ironwood). TensorFlow is not.

• TPU7x (Ironwood) specifications
  Per-chip and per-pod specifications for TPU7x (Ironwood).

• TPU7x (Ironwood) supported configurations
  The slice shapes you can request for TPU7x (Ironwood), and how many chips and hosts each uses.

• TPU7x (Ironwood) VM types
  The VM types available for TPU7x (Ironwood) and the host resources each one provides.

• TPU7x (Ironwood) memory hierarchy
  How high-bandwidth memory (HBM) and host memory are organized on a TPU7x (Ironwood) chip.

Related information

TPU7x (Ironwood) specifications

TPU7x (Ironwood) supported configurations

TPU7x (Ironwood) VM types

Create a TPU VM