# Google introduces two new TPU system configurations

> New configurations target confidential workloads and memory demands.

By BIGCHANGE.AI Editorial

Published: 2026-11-10T06:41:39.470Z
Updated: 2026-11-10T06:41:39.470Z
Canonical: https://bigchange.ai/blog/google-tpu-8tx-8i-max-systems

![Illustrative accelerator board in a shallow tray on a bench.](https://bigchange.ai/api/media/file/system-board-v2.png)
AI-generated illustration for BIG CHANGE

Google [announced](https://cloud.google.com/blog/topics/ai-infrastructure/tpu-8tx-and-tpu-8i-max) two TPU systems on November 9: 8tx for confidential computing and 8i Max for memory-intensive AI. Both reuse existing eighth-generation silicon. Its [product page](https://cloud.google.com/tpu) lists them as coming soon, without dates or prices.

## Design and limits

Google says 8tx uses TPU 8t silicon with fifth-generation AMD EPYC hosts, providing hardware-level protection for data and model weights during processing. It targets confidential inference and small-scale training. The system also uses the low-latency Boardfly network.

Protecting information while it is being processed is different from encrypting stored data or network traffic. Google’s separate [Confidential VM documentation](https://docs.cloud.google.com/compute/docs/about-confidential-vm) describes encryption in memory and security controls for those virtual machines. It does not establish that every VM feature or configuration is available on 8tx. Customers evaluating sensitive workloads need the actual system’s supported controls and threat model before relying on it.

For 8i Max, Google places two TPU 8i chips in one package. It claims double the compute, SRAM and high-bandwidth memory bandwidth of 8i. Those are configuration comparisons, not independent measurements of application speed. The announcement does not specify the new package’s total memory capacity.

## Why memory matters

An AI system must move data between processors and memory while generating an answer. Long inputs and multi-step workloads can make that movement a constraint, so bandwidth and nearby memory matter alongside arithmetic capacity. Performance still depends on the model, input, software and workload; a doubled hardware specification is not a promise of twice the throughput.

Google’s April [technical discussion](https://cloud.google.com/blog/products/compute/tpu-8t-and-tpu-8i-technical-deep-dive) describes the original 8i’s Boardfly interconnect and its 288 GB of high-bandwidth memory per chip. It explains how the design reduces communication distance among chips.

Teams should compare security and performance against each workload’s requirements. Pricing, supported software, capacity and workload-specific tests remain necessary to assess either choice.

## Sources

- [Google TPU systems announcement](https://cloud.google.com/blog/topics/ai-infrastructure/tpu-8tx-and-tpu-8i-max) — System designs and claims.
- [Google Cloud TPU product page](https://cloud.google.com/tpu) — Current product status.
- [TPU 8t and TPU 8i technical discussion](https://cloud.google.com/blog/products/compute/tpu-8t-and-tpu-8i-technical-deep-dive) — Earlier chip architecture.
- [Confidential VM documentation](https://docs.cloud.google.com/compute/docs/about-confidential-vm) — General security context.