Google has unveiled two more TPU systems, pairing one with AMD processors and cramming two accelerator chips into the other package.
According to Reuters, the 8tx and 8i Max expand Google’s attempt to sell customers an alternative to Nvidia’s AI hardware.
The company is adapting existing eighth-generation silicon to different jobs. TPU 8tx combines the silicon used in TPU 8t with fifth-generation AMD EPYC host processors. Google has changed the surrounding system to support confidential computing.
Target customers include governments and regulated businesses that need to protect sensitive information while models process it.
Google says its hardware isolates data and model weights during use. Ordinary encryption for storage and network traffic leaves that processing stage as a separate problem.
The system uses Google’s low-latency Boardfly network. Google claims latency can be reduced by up to half against TPU 8t, with negligible overhead from the confidentiality protections.
That is a manufacturer comparison, and says little about how a particular customer’s application will behave. Buyers still need measurements for their own models.
The 8i Max tackles memory constraints by putting two TPU 8i chips into one package. Google says compute, SRAM capacity and HBM bandwidth each double.
Large context windows can leave processors waiting for data, no matter how impressive their arithmetic specifications look. Additional bandwidth is useful when it keeps those units occupied.
Google says the package increases compute density within the same rack footprint. Power and cooling requirements still deserve scrutiny alongside any boast about floor space.
Reuters reports Google Cloud’s quarterly revenue more than doubled to approximately €27.44 billion, with TPU sales making a meaningful contribution. Google is increasingly treating its chips as a business. The company’s launch announcement provides no firm availability date or pricing for the two new configurations.
