Google has broadened its portfolio of in-house developed AI chips with the introduction of two new Tensor Processing Units (TPUs): the TPU 8tx and TPU 8i Max. These additions, unveiled on Monday, are designed to support new use cases, including confidential model training and increased memory capacity for large AI workloads.
The TPU 8tx, developed with AMD, is based on the existing TPU 8t chip. Google states it is specifically engineered for confidentially training and running AI models, integrating security directly into the hardware layer to protect sensitive data and model weights. This chip also features a custom inferencing network, allowing it to perform reasoning alongside training, which Google suggests could be useful for training models on proprietary data without moving it between devices.
Meanwhile, the TPU 8i Max combines two inference-focused TPU 8i chips, effectively doubling their memory capacity, processing power, and bandwidth within a single package. This configuration allows AI models to handle a wider range of information during queries.
These new chips are part of Google’s eighth generation of TPUs, first introduced in April this year, and are intended to reduce the company's dependence on chips from other manufacturers for its cloud computing business. The launch aligns with a broader trend among hyperscalers and AI labs, such as Microsoft, Amazon, OpenAI, and Anthropic, who are developing their own AI-specific semiconductors to lessen reliance on Nvidia.