Google expands its AI hardware portfolio with two new TPU chips

Google has added the TPU 8tx and TPU 8i Max models, designed for operations requiring confidential inference and high memory capacity, to its family of Tensor Processing Units (TPUs) developed for AI workloads.

Based on the eighth-generation chip architecture introduced in April, the new models stand out as part of Google's strategy to reduce its reliance on external suppliers like Nvidia in its cloud computing infrastructure. Designed in collaboration with AMD, the TPU 8tx provides security at the hardware layer, allowing sensitive data and model weights to be isolated during active use. This chip also features a specialized inference network capable of running reasoning processes during training.

The TPU 8i Max, on the other hand, was created by combining two previously released TPU 8i chips into a single package. This structure doubles memory capacity, processing power, and bandwidth, enabling AI models to process larger data sets during querying. Other major players in the sector, such as Microsoft, Amazon, OpenAI, and Anthropic, likewise aim to increase their hardware independence by developing their own AI-focused semiconductors.