Google unveils two new AI TPU chips for cloud services — TechCrunch

Google has unveiled two new versions of its proprietary tensor processing units for artificial intelligence tasks — TPU 8tx and TPU 8i Max. They complement the eighth generation of TPU chips, which the company announced in April, and are intended for confidential work with models and heavy workloads, respectively, TechCrunch reports.

Confidential work with models

TPU 8tx, developed with AMD based on TPU 8t, is intended for confidential training and running of AI models. According to Google, protection is integrated directly at the hardware level: sensitive data and model weights are meant to remain protected and isolated even during active use.

The chip also received a dedicated inference network, allowing reasoning tasks to be performed in parallel with model training. Google said this could make it possible to distill and train smaller models during operation without moving a company’s proprietary data between devices.

More current news is available on the UA.News Telegram channel Telegram.

More memory in one package

TPU 8i Max effectively combines two inference-oriented TPU 8i chips, introduced in April, in one package. This doubles the memory capacity, computing power, and bandwidth compared with a single TPU 8i.

According to Google, this will allow AI models to process a broader volume of information during queries. Expanding its lineup of proprietary processors is part of the company’s efforts to reduce its cloud business’s dependence on chips from other manufacturers.