Google launches TPU 8tx, TPU 8i Max

Google launched two new TPUs to bring its launch cadence to four new AI chips this year.

The company launched TPU 8tx and TPU 8i Max. At Google Cloud Next, Google launched TPU 8t for training and TPU 8i for inference.

Google is a chipmaker in its own right and TPUs are a big reason for Google Cloud's revenue surge.

In a blog post, Google said TPU 8i Max is designed for massive context windows and high compute density per rack workloads. TPU 8tx is designed for secure and confidential inference.

Google is ramping its flavors of TPUs due to a wide range of use cases for AI as agents scale. For instance, Google customers are increasingly asking for confidential inference and specialized encryption found in TPU 8tx.

According to the company, the TPU lineup is designed to be a building block that can adapt as AI software changes. Key points about the new chips include:

• TPU 8tx is optimized for confidential inference but also can be used for small-scale training.
• TPU 8tx is paired with AMD's EPYC CPU.
• TPU 8tx security is built into the hardware layer and keeps sensitive data and model weights isolated when active.
• TPU 8tx has 2x lower latency than TPU 8t.
• TPI 8i Max is built for the memory wall and has double the SRAM and HBM bandwidth of TPU 8i.