Google unveils Ironwood, its seventh-generation TPU designed for inference

2025-04-09 · Publisher : Google Cloud · Domain : Cloud & DevOps

At Google Cloud Next 2025, Google introduces Ironwood, its seventh-generation TPU, which it describes as the first designed for inference. A pod of 9,216 chips reaches 42.5 exaflops, with 192 GB of HBM memory per chip.

### Background On 9 April 2025, at Google Cloud Next, Google introduces Ironwood, its seventh-generation TPU (Tensor Processing Unit). Google describes it as its most performant and scalable AI accelerator to date, and as the first TPU designed specifically for inference, that is, for serving models rather than training them. Availability is announced for later in 2025. ### Key facts - Per chip: 4,614 teraflops of peak compute, 192 GB of HBM memory and 7.37 TB/s of memory bandwidth. - Inter-chip interconnect (ICI) of 1.2 Tbps bidirectional. - Two configurations: 256 chips or 9,216 chips; the latter reaches 42.5 exaflops in total. - Announced efficiency: twice as good per watt as Trillium (sixth generation) and nearly 30 times better than the first Cloud TPU of 2018. - Memory capacity multiplied by 6 and memory bandwidth by 4.5 compared with Trillium. - Google's Pathways runtime coordinates distributed computing across very large sets of chips. - Later clarification from Google (6 November 2025): up to 10x the peak performance of TPU v5p and more than 4x the per-chip performance of TPU v6e (Trillium) for training and inference. ### Why it matters Inference cost weighs more and more in AI spending as reasoning models are deployed. Ironwood targets that precise point and aims to lower the cost per request against GPUs. For customers, the benefit will depend on software portability (JAX, PyTorch) and on real pricing, which Google did not detail at the announcement. ### Takeaway The seventh-generation TPU targets inference: the AI accelerator specializes by use.

Tags : Google Cloud, Puces, Hyperscaler, IA, Conférence