Google Expands TPU Lineup
Google is advancing its custom chipmaking efforts to reduce reliance on external suppliers for its cloud computing infrastructure. The company unveiled two new variants of its tensor processing units, which are custom chips designed for artificial intelligence workloads, expanding its hardware portfolio to support confidential inference and larger memory capacities.
The new hardware, designated as the TPU 8tx and TPU 8i Max, are additions to the eighth generation of processors introduced earlier in the year, building upon previous capabilities for inference capacity and large-scale model training requirements.
Confidential Inference and Security
Developed in collaboration with AMD and based on the TPU 8t architecture, the TPU 8tx is engineered specifically for confidential AI model training and execution. The hardware integrates security directly into the physical layer to keep sensitive data and model weights isolated and protected during active usage.
The processor also features a custom inferencing network enabling reasoning operations alongside training. This allows enterprises to distill and train smaller models dynamically using proprietary data without requiring data transfers across external devices.
High-Capacity Memory Design
The TPU 8i Max integrates two inference-focused TPU 8i chips into a single package, effectively doubling processing power, bandwidth, and memory capacity. This unified design enables AI models to process broader ranges of information concurrently during complex queries.
Industry Context
The release arrives as major hyperscale providers and artificial intelligence laboratories develop proprietary semiconductors and software ecosystems to decrease dependence on dominant market leaders like Nvidia.




