Google Launches Two In-House Chips: TPU 8TX and TPU 8I Max
Google announced the launch of two new in-house tensor processors, TPU 8TX and TPU 8I Max, expanding the eighth-generation TPU family to four systems within a year. TPU 8TX focuses on full-stack confidential inference; it is the first TPU system to integrate hardware confidential computing features, and it introduces the Boardfly network topology, reducing the inter-chip communication diameter by more than 50% and halving inference latency compared with the previous generation. TPU 8I Max uses a dual-chip package, integrating two 8I chips within a single module, providing 17.2 TB/s of memory bandwidth, 768 MB of on-chip SRAM and 384GB of HBM; within the same rack space and power consumption, its compute density and data throughput are twice those of TPU 8I. Both chips and the acceleration software will be available on Google Cloud.
Why read this
AI score 62
The original text presents the positioning and key specifications of the two TPUs, allowing readers to understand the division of labor of Google's in-house compute in confidential inference and large-model inference.
Topics
Body · AI translation
Google announced yesterday (the 9th) the launch of two new self-developed tensor processors (TPUs), "TPU 8TX" and "TPU 8I Max," as it moves to fully expand its pursuit of AI business opportunities.
Industry analysts note that Google's eighth-generation TPU family, spanning from cutting-edge process technology to system-level packaging, is currently tied to the TSMC ecosystem. As Google continues to expand its related product line lineup and gains adoption from more major companies, TSMC stands to benefit as well.
Mark Lohmeyer, Vice President and General Manager of Google Cloud AI and Compute Infrastructure, said that Google's past cadence of launching one AI chip every two years or every year will accelerate, and it has now reached the level of "launching four dedicated systems in a single year," responding to market demand with extremely rapid hardware R&D iteration.
Google only launched the TPU 8T, aimed at hyperscale training, and the TPU 8I, focused on low-latency inference, in the first half of this year. Now it is once again expanding the eighth-generation TPU family with the TPU 8TX and TPU 8I Max, reaching the milestone of unveiling four TPU systems within one year.
Industry analysts note that the focus of AI technology development is shifting from "Systems of Record" to "Systems of Action" that can link with enterprise ERP and databases and take autonomous action, causing underlying compute loads to surge exponentially.
The two new Google chips each have their own strategic positioning. The TPU 8TX focuses on full-link confidential inference and is the first TPU system to integrate hardware confidential computing features, deeply integrated with the AMD x86 processor architecture responsible for system control. The chip uses end-to-end PCIe link hardware encryption and identity authentication to ensure that, during inference in finance, healthcare, or strictly regulated industries, model weights and private data remain isolated and protected; it also introduces a new network topology architecture, Boardfly, which reduces the communication diameter between chips by more than 50%, cutting inference latency by more than half compared with the previous generation.
The TPU 8I Max focuses on carrying massive documents, code repositories, and parallel agent workflow designs. It adopts a dual-chip packaging architecture, integrating two 8I chips within a single module, providing up to 17.2 TB/s of memory bandwidth (a twofold increase), 768 MB of on-chip SRAM (a twofold increase), and 384GB of HBM (a 33% increase), delivering twice the compute density and data throughput of the TPU 8I under the same rack space and power consumption conditions.
Google emphasizes that future compute architectures will no longer be a simple contest of the number of accelerators, but a comprehensive integrated design of silicon chips, network interconnection, memory bandwidth, and scheduling software. The new chips and acceleration software will all launch on Google Cloud, building next-generation AI infrastructure for global enterprises that combines top-tier performance, elastic scheduling, and high compliance standards.