Google has announced the expansion of its eighth-generation Tensor Processing Unit (TPU) family for AI, launching two new systems: TPU 8tx and TPU 8i Max. This announcement strengthens AI infrastructure by focusing on the requirements of confidential computing and handling long-context workloads.
TPU 8tx System for Confidential Computing
The TPU 8tx system is specifically designed to meet the needs of organizations handling sensitive data subject to strict regulatory requirements. The system is based on existing TPU 8t chips, integrating fifth-generation AMD EPYC server processors for task management and data orchestration. This configuration allows users to run AI models within Confidential Computing environments, providing robust protection for model weights and data during processing, storage, and transit, thanks to Secure Encrypted Virtualization (SEV) technologies.
Google also noted that the integration of the low-latency Boardfly network has helped reduce latency by half compared to the standard TPU 8t system, making it suitable for fine-tuning and reinforcement learning operations.
TPU 8i Max System for Long Contexts
The TPU 8i Max system targets workloads requiring massive memory and high processing power, such as autonomous agent applications that process large software repositories or extended visual content. This system is based on clustering two TPU 8i units into a single package, which doubles the computational power, SRAM memory, and HBM memory bandwidth, thereby increasing compute density within data centers.
What This Expansion Means for Users
This announcement represents Google's strategy to diversify its technical options without needing to design entirely new chips, as the company focuses on optimizing the integration of compute, memory, and networking. With the addition of these two systems, the eighth-generation family now includes four options covering large-scale training, fast inference, confidential computing, and long-context processing. Google reaffirms its commitment to supporting open-source frameworks such as PyTorch and the TorchTPU project to facilitate the adoption of these systems by developers.
It is worth noting that Google has not yet disclosed pricing details or timelines for the commercial availability of these systems, nor have any specific projects been announced in the Arab region.