
Google adds TPU variants for confidential inference and long-context workloads
The AMW Read
The variants broaden a known accelerator lineup, while system sales and software support affect competition in AI infrastructure.
Google adds TPU variants for confidential inference and long-context workloads
Google introduced two TPU variants aimed at confidential inference and long-context workloads, according to a 36Kr report. The additions bring its eighth-generation TPU lineup to four configurations, alongside systems for training and real-time inference. The report also says Google began delivering complete TPU systems to selected customers for use in their own data centers.
The move expands Google's role in AI infrastructure from providing cloud access to its custom chips to supplying complete systems outside its data centers. The report says those systems include compute, storage, networking hardware and software. This gives some customers another way to obtain AI compute, while Google's work with Meta on the open-source TorchTPU project targets a practical barrier: adapting PyTorch models to run on TPUs.
For builders evaluating accelerators, the relevant test is performance per dollar on their own models, including the effort required to migrate and maintain software. The report cites tests in which seventh-generation Ironwood TPUs outperformed Nvidia systems on that measure in some cases, but also describes Nvidia regaining an advantage on a model after optimization. Those results do not establish a universal winner across workloads.



