Google has announced two AI chip specialists for its eighth-generation Tensor Processing Unit family, according to SiliconANGLE’s November 9 report. TPU 8i Max targets long-context processing, while TPU 8tx is designed for confidential inference—the work of running a model. Source
AMD separately announced that fifth-generation EPYC processors will power the 8tx’s host infrastructure—the CPU side of the system. Source
8i Max: more memory, same rack footprint
According to SiliconANGLE, Google says 8i Max delivers twice the compute, twice the SRAM—on-chip memory—and twice the high-bandwidth memory of TPU 8i. The company also says it fits twice the silicon into the same rack footprint. More silicon, no bigger office. Source
A context window is how much information an AI model can consider while producing an answer. Longer documents, conversation histories and collections of code give the model more material to work with, but also increase demands on memory and data movement. Source
Those specifications don’t establish that an application will run twice as fast. Source
8tx: the sensitive-data assignment
The 8tx targets regulated businesses and governments that want to run models or AI agents alongside sensitive data. SiliconANGLE reports that it uses the TPU 8t architecture with AMD host CPUs and incorporates a hardware secure enclave intended to protect and isolate sensitive data and model weights. The security concern here is information being processed, not just files sitting in storage. Source
Google also says the system combines training-grade compute with fast inference. According to SiliconANGLE, the announced capabilities include fine-tuning models with private data, distilling larger models into smaller ones and training agents through reinforcement learning, all on a single confidential system without moving data between separate clusters.
