The AI computing power competition has escalated, with Google releasing the next-generation Ironwood TPU architecture, which boasts a 16-fold increase in performance, achieving a single-chip computing power of 4614 TFLOPs

The arms race for AI infrastructure is escalating at an unprecedented pace. Google's latest next-generation Tensor Processing Unit (TPU) platform, Ironwood, has once again raised the bar in this competition with its astonishing performance leap.

According to information disclosed by Google at the Hot Chips 2025 conference, its seventh-generation TPU architecture, Ironwood, has achieved exponential growth in core performance, with a peak computing power of up to 4614 TFLOPs for a single Ironwood chip. Compared to the TPU v4 launched by Google in 2022, Ironwood's single-chip computing power has increased by more than 16 times; even compared to the TPU v5p released last year, it has grown nearly 10 times.

The release of Ironwood is not just an innovation in a single chip, but a complete system-level solution designed for extreme scalability. Google also announced the racks, network interconnections, and cooling systems built around this chip, showcasing its full-stack capability to transform cutting-edge computing power into large-scale, high-efficiency productivity.

Performance Leap: Single Chip Computing Power Increased by Over 16 Times

The data released by Google clearly demonstrates the evolution of its TPU platform's performance. Specifically, Ironwood's single-chip peak computing power reaches 4614 TFLOPs and is equipped with 192 GB of high-bandwidth memory (HBM), with a bandwidth of up to 7.4 TB/s. In comparison, the TPU v4 released in 2022 had a single-chip computing power of 275 TFLOPs, equipped with 32 GB HBM and a bandwidth of 1.2 TB/s. The TPU v5p launched in 2023 had a single-chip computing power of 459 TFLOPs, equipped with 95 GB HBM and a bandwidth of 2.8 TB/s.

From TPU v4 to Ironwood, Google has increased single-chip computing power by more than an order of magnitude in just a few years, reflecting the explosive growth in computational demands of AI models and the efforts of chip designers to meet this demand.

At the Superpod level, an Ironwood Superpod will contain 9216 chips, further expanding the scale compared to previous generations.

System Architecture: Scaling from Chips to Supercomputing Clusters

Powerful chip performance must rely on precise system design to be fully realized. Google detailed the modular and scalable architecture of Ironwood from chips to racks to clusters. At the core of the system is the Ironwood SoC (System on Chip), with four such chips integrated onto a single Ironwood PCBA motherboard.

Subsequently, 16 PCBA motherboards are stacked like trays to form an Ironwood TPU rack containing 64 chips. Inside the rack, Google has adopted a 4x4x4 3D Torus network topology that has been used in at least three generations of products, forming a logical computing unit. To achieve larger-scale expansion, Google employs its proprietary Inter-Chip Interconnect (ICI) technology, connecting multiple racks into a Superpod through a mix of PCB traces, copper cables, and fiber optic links.

According to reports, the system can connect up to 43 computing units (each unit containing 64 chips), forming a massive cluster with 1.8 Petabytes of network bandwidth.

Behind the astonishing computing power lies significant energy consumption and heat dissipation challenges. Data shows that a fully loaded Ironwood rack can consume over 100 kilowatts, placing stringent demands on the power supply and cooling systems of data centers. To address this challenge, Google has equipped the Ironwood rack with an efficient liquid cooling system.

This system includes a CBU rack for coolant distribution and a drip tray installed at the top of the rack to monitor any potential liquid leaks. In terms of power supply, the rack uses a 416-volt AC input, converted by a rectifier to provide DC power to the system