Skip to main content

Google Cloud TPU v7 (Ironwood)

Overview​

Google TPU v7 (codenamed Ironwood) is the latest-generation TPU, launched in 2025 and purpose-built for the era of inference. Per-chip BF16 compute reaches 2,307 TFLOPS, with 192GB HBM (on par with NVIDIA H200 / MI300X) and 7,380 GB/s bandwidth. Ironwood is among the first inference-first TPUs, delivering 4,614 TFLOPS FP8 per chip.

Core Specifications​

ItemSpecification
ArchitectureTPU v7 (Ironwood)
BF16 Compute (per chip)2,307 TFLOPS
FP8 Compute (per chip)4,614 TFLOPS
HBM Capacity192 GB
HBM Bandwidth7,380 GB/s
ICI Interconnect Bandwidth1,200 GB/s (bidirectional)
DCN Bandwidth100 Gbps
TensorCores2/chip
SparseCores4/chip
Pod Size9,216 chips
Topology3D Torus

TPU Generations Comparison​

Metricv4v5pv6e (Trillium)v7 (Ironwood)
BF16 Compute275 TFLOPS459 TFLOPS918 TFLOPS2,307 TFLOPS
FP8 ComputeN/A459 TFLOPS918 TFLOPS4,614 TFLOPS
HBM Capacity32 GB95 GB32 GB192 GB
HBM Bandwidth1,200 GB/s2,575 GB/s1,638 GB/s7,380 GB/s
Pod Size4,0968,9602569,216

Ironwood vs H200 / MI300X​

MetricTPU v7H200MI300X
Memory192 GB141 GB192 GB
Bandwidth7,380 GB/s4,800 GB/s5,300 GB/s
FP8 Compute4,614 TFLOPS3,958 TFLOPS2,614 TFLOPS

Key advantage: TPU v7 leads in memory bandwidth and FP8 compute.

Vendor Information​

ItemDetails
ManufacturerGoogle LLC
Official Websitehttps://cloud.google.com/tpu
Product Pagehttps://cloud.google.com/tpu/docs/tpu7x
Release2025
AvailabilityGoogle Cloud only

Key Features​

  • Inference-first: Optimized for inference, unlike prior generations that focused on training
  • Massive memory capacity: 192GB enables 70B+ models on a single chip
  • FP8 at 2× BF16: Modern inference compute
  • 3D Torus 9,216-chip Pod

Use Cases​

  • LLM inference (Gemini 3 / 4)
  • Large-scale MoE models
  • Multimodal AI inference
  • Embedding-intensive applications