At its Cloud Next conference this week, Google revealed Ironwood, the latest generation in its line of TPU AI accelerator chips.
Ironwood, which marks Google’s seventh-generation TPU, is notably the company’s first processor specifically designed and optimized for inference tasks—running existing artificial intelligence models rather than training new ones. According to Google, this newest chip promises significantly enhanced capabilities across both power efficiency and processing capacity.
Google Cloud VP Amin Vahdat described Ironwood as “our most powerful, capable, and energy-efficient TPU yet,” highlighting that it is purpose-built to accelerate large-scale inference workloads.
Ironwood will debut sometime later this year, becoming available to Google Cloud customers in two distinct configurations: clusters of 256 chips and massive clusters comprising as many as 9,216 chips. Google’s internal benchmarks indicate Ironwood can achieve peak computing performance of up to 4,614 TFLOPs per chip and includes 192GB of dedicated RAM boasting a data bandwidth of nearly 7.4 terabits per second.
A key innovation featured in Ironwood is SparseCore, an enhanced specialized processing core designed specifically to handle sparse data. This type of workload is common in complex recommendation systems and advanced ranking algorithms, such as those used by online platforms to suggest products or content. The new architecture substantially reduces data movement and latency, offering significant improvements in speed and overall efficiency.
Google also announced it plans to integrate Ironwood processors into its broader initiative—the AI Hypercomputer—a modular computing cluster within the Google Cloud infrastructure, intended to meet escalating demand for powerful, scalable AI workloads.
Ironwood’s introduction comes at a moment when the AI chip market has become increasingly competitive. While Nvidia continues to hold a commanding position, other tech giants like Amazon, with its Inferentia, Trainium, and Graviton chips, and Microsoft, with their custom-built Azure-hosted Cobalt processors, are ramping up their investments in specialized AI hardware.
“Ironwood represents a unique breakthrough in the age of inference,” said Vahdat. “Its significant advancements in computation, memory capacity, networking efficiency, and reliability set a new benchmark for AI acceleration.”
