Unveiled at a cloud conference this week, Alphabet’s seventh-gen AI processor, Ironwood, is made for inference workloads.

Google VP Amin Vahdat said: “This chip is designed to run AI applications, not just train them. The relative importance of inference is going up significantly.”

Unlike Nvidia’s kit, Google’s Tensor Processing Units (TPUs) are bespoke gear for internal devs and cloud punters.

Earlier chip generations split duties between model training and cost-cutting inference. Ironwood merges those camps, adding more memory to handle large AI deployments like Google’s Gemini models. According to Vahdat, Ironwood delivers twice the performance per watt compared to last year’s Trillium chip.

Google said each Ironwood chip has a peak compute of 4,614 teraflops (TFLOP), which is considerably higher throughput than its predecessor, Trillium, which was unveiled in May 2024. The tech giant also plans to make these chipsets available as clusters to maximise the processing power for higher-end AI workflows.

Ironwood can be scaled to a cluster of 9,216 liquid-cooled chips linked with an Inter-Chip Interconnect (ICI) network. The chipset is also one of the new components of Google Cloud AI Hypercomputer architecture. Developers on Google Cloud can access Ironwood in two sizes: a 256—chip configuration and a 9,216-chip configuration.

Ironwood chipsets can generate up to 42.5 Exaflops of computing power at its most expansive cluster.

Google claimed its throughput is more than 24x of the compute generated by the world's largest supercomputer, El Capitan, which offers 1.7 Exaflops per pod. Ironwood TPUs come with expanded memory, with each chipset offering 192GB, sextuple of what Trillium was equipped with. The memory bandwidth has also been increased to 7.2Tbps.

Google has kept quiet about who’s fabbing Ironwood. There is no tip whether it’s TSMC, Samsung, or someone else etching their secret sauce into silicon.