The TPU 8tx protects data in use at the hardware level. The dual-chip TPU 8i Max targets workloads that need to move more data through memory.
Google introduced two configurations of its eighth-generation tensor processing units on Monday, extending the range of AI hardware it offers through its cloud business. The TPU 8tx is designed for work involving sensitive information, while the TPU 8i Max pairs two processors for tasks that need more computing power and memory bandwidth.
The new systems build on Google’s TPU 8t and 8i, announced in April for large-scale training and low-latency reasoning respectively. Google said the underlying silicon was already in use and described the latest products as new combinations of processors, memory, networking and security features.
Hardware protection for data in use
The 8tx combines 8t silicon with fifth-generation AMD EPYC host processors. Google says it can isolate data and model weights during processing, when they would otherwise need to be exposed to the system running the computation. The company is targeting governments and regulated industries that want to run proprietary models or adapt them using internal data.
Google also said the 8tx uses a low-latency interconnect from its reasoning systems. The company says this can cut latency by up to half compared with the TPU 8t. Performance and security in a deployment will also depend on the workload and how customers configure the surrounding system.
Two chips for memory-intensive workloads
The 8i Max places two 8i chips in one package. Google says the package doubles compute capacity, on-chip SRAM and high-bandwidth memory throughput relative to one 8i. That extra bandwidth is intended for jobs such as analysing long video inputs or running software assistants across large codebases.
Both systems expand Google Cloud’s options for customers with different requirements. Reuters reported that TPU sales helped drive the cloud unit’s September-quarter revenue, which more than doubled to $30.71 billion.
