Google Details Its Most Powerful AI Processor, Ironwood

Tech giant Google detailed its Ironwood TPU Superpod solution at Hot Chips 2025. The new system delivers a performance increase of over 16 times compared to the previous generation.
At the Hot Chips 2025 event, Google shared detailed information about its next-generation artificial intelligence accelerator, “Ironwood” TPU Superpod, which was introduced in recent months. This system, built on the company’s 7th generation TPU architecture, offers a massive performance boost compared to previous generations.

The Ironwood TPU comes with 9216 chips in each pod, 192 GB of HBM memory per chip, and a processing power of 4614 TFLOPs. These figures correspond to more than 16 times the performance offered by the TPU v4 introduced in 2022. This giant structure can achieve a total computing power of 42.5 Exaflops.
The company claims that Ironwood is 24 times faster than similar system segments of El Capitan, known as the world’s most powerful supercomputer. Google makes this comparison based on FP8, but El Capitan, which offers 1.7 Exaflops, does not support this.
Ironwood Solution Detailed

The TPU v4 reached 275 TFLOPs with 4096 chips and 32 GB of HBM memory per chip, while the TPU v5p, which arrived a year later, offered 8960 chips, 95 GB of memory, and a capacity of 459 TFLOPs. Ironwood raises this bar much higher.
At the heart of Ironwood is the new generation Ironwood SoC. Four chips are placed on each motherboard, and 16 motherboards are combined to form a rack. This structure is designed as a modular block of 64 chips. Google connects these blocks using a network architecture called 3D Torus. This 4x4x4 structure provides low latency and high flexibility in data communication.

The system also uses Google’s InterChip Interconnect (ICI) technology. This enables scalability up to 43 blocks, achieving a network capacity of 1.8 petabytes. An Ironwood Superpod contains a total of 144 racks. Additionally, optical switch cases and liquid cooling units are integral parts of this structure. Google supports the entire system with hybrid connection technologies. This means copper connections on the motherboard are combined with fiber optic lines to provide both flexibility and high bandwidth.
On the energy front, each rack operates with a custom power distribution that converts a 416-volt AC input to a DC output. The entire system is liquid-cooled, and a single rack can consume over 100 kW of power.
You Might Also Like;
- AI Tools Are Coming to Opera Android: Here Are the Innovations
- Top 6 Essential AI APIs for Developers to Integrate Into Your Projects in 2026
- 10 Best AI Tools on Your Phone: Top Mobile Apps for iOS & Android (2026)
Follow us on TWITTER (X) and be instantly informed about the latest developments…









