Google TPU 8tx Pairs AMD EPYC Turin Hosts with Confidential Computing (www.servethehome.com)

Google has announced updates to its TPU portfolio with the introduction of the TPU 8tx and TPU 8i Max systems, expanding its eighth-generation offerings to four configurations in a single year. The TPU 8tx is notable for its integration of AMD's 5th Gen EPYC "Turin" hosts, enhancing confidential computing capabilities for regulated enterprises and government applications. This system keeps data encrypted even during processing by utilizing a trusted execution environment, addressing security concerns during inference tasks. Significantly, it maintains low latency through its Boardfly networking topology, which reduces the maximum chip hops compared to previous designs. In addition to the TPU 8tx, the TPU 8i Max combines two TPU 8i chips to effectively tackle memory bandwidth limitations in large-context workloads. This dual-chip configuration doubles the compute power and memory resources, delivering 384 GB of high-bandwidth memory and a bandwidth increase to 17.2 TB/s. Google’s tailored approach to these TPU systems illustrates its commitment to optimizing performance for specific AI tasks, distinguishing between confidential computing and inference needs. This development is a strategic win for AMD as it secures a socket in a major hyperscaler environment, showcasing its capabilities in confidential computing alongside Google's cutting-edge AI hardware.