Google announces TPU 8i Max dual‑chip and confidential TPU 8tx

Google added two new variants to its eighth‑generation Tensor Processing Unit (TPU) family, aiming to address the growing demand for longer context windows and privacy‑preserving inference.

What are the TPU 8i Max and TPU 8tx chips?

According to the MarkTechPost report, the TPU 8i Max bundles two TPU 8i chips inside a single package. By pairing the chips, the Max version can keep a larger amount of intermediate data in fast on‑chip memory, which is useful for models that need to process long sequences or maintain extensive context across many tokens. The TPU 8tx, on the other hand, introduces a hardware‑based confidential inference mode. This mode aims to keep user data encrypted while it is being processed, reducing the surface for data leakage during inference.

How do they differ from the existing TPU 8i?

The standard TPU 8i, released earlier in the eighth‑generation line, already supports high‑throughput matrix multiplication for modern large language models. The Max version doubles the compute capacity by physically placing two of those chips together, effectively offering twice the on‑chip memory bandwidth without a proportional increase in power draw. The 8tx does not change raw compute performance; instead, it adds a secure enclave that executes inference inside a protected environment, a feature that was previously only available through software‑level encryption layers.

Which workloads are the new chips designed for?

Long‑context agents—such as conversational assistants that need to remember the entire dialogue history, or retrieval‑augmented generation systems that scan large document collections—benefit from the expanded memory footprint of the 8i Max. By keeping more tokens in fast memory, these agents can avoid costly off‑chip data swaps, leading to lower latency and higher throughput.

The 8tx is targeted at industries where data confidentiality is paramount, including finance, healthcare, and regulated enterprise AI services. Its hardware confidential inference capability means that sensitive inputs (for example, patient records or credit‑card data) can be processed without ever being exposed in clear text to the host system.

When will the chips be available, and who can obtain them?

Bloomberg, cited in the MarkTechPost article, indicates that both chips are expected to ship in 2027. Access will likely follow Google’s existing TPU Cloud offering, where developers and enterprises can reserve accelerator instances on Google Cloud. Early‑access programs may be opened to select research partners before the broader commercial rollout.

How can Grafy users experiment with these TPUs?

While Grafy does not yet provide direct access to Google’s TPU hardware, developers can still prepare their models for TPU deployment using the Grafy Code studio. The coding environment lets you write, test, and profile TensorFlow or JAX workloads, which are the primary frameworks supported by Google TPUs. When the hardware becomes publicly available, you can migrate the same code to a TPU‑enabled Cloud instance with minimal changes.

For creators interested in comparing model outputs across different back‑ends, the Grafy Chat interface lets you run the same prompt against a variety of hosted models side‑by‑side. Although the chat does not run on TPUs, you can use it to benchmark performance and quality before moving to a TPU‑accelerated deployment. Both apps are reachable from the Grafy dashboard:

• Grafy Code
• Grafy Chat

What does this mean for the broader AI ecosystem?

The dual‑chip design of the 8i Max signals that hardware vendors are acknowledging the “long‑context” problem as a first‑order concern for next‑generation AI services. By providing more on‑chip memory, Google is reducing the reliance on external storage, which can be a bottleneck for large language models that need to retain hundreds of kilobytes of context.

Meanwhile, the confidential inference feature of the 8tx reflects a growing regulatory and market pressure for privacy‑by‑design AI. If hardware can guarantee that data never leaves an encrypted enclave, developers can meet compliance requirements more easily, potentially unlocking new use cases in regulated sectors.

Together, the two chips illustrate a split strategy: one focused on raw performance for richer interactions, the other on security for sensitive workloads. As more cloud providers adopt similar capabilities, we can expect a wave of applications that combine deep, context‑aware reasoning with strong privacy guarantees.

---

Source: MarkTechPost

Frequently asked questions

What is the main advantage of the TPU 8i Max over a single TPU 8i?

The 8i Max combines two TPU 8i chips in one package, effectively doubling on‑chip memory and bandwidth. This allows models that need to retain longer context windows to run with lower latency and fewer off‑chip memory accesses, which is especially useful for large‑scale language models and retrieval‑augmented systems.

How does hardware confidential inference differ from software encryption?

Hardware confidential inference runs the model inside a secure enclave built into the TPU itself. Data remains encrypted while it is stored, transferred, and processed, reducing the attack surface compared to software‑only encryption, which must rely on the host operating system and can be vulnerable to memory‑scraping attacks.

When can developers start using the new TPUs on Google Cloud?

Bloomberg reports that both the TPU 8i Max and TPU 8tx are slated for a 2027 release. Google typically rolls out new accelerator types first to a limited set of research partners before opening them to the broader Cloud Marketplace, so early access may be possible for select collaborators before the general availability date.

---

Source: MarkTechPost — Google Unveils TPU 8i Max and TPU 8tx: A Dual-Chip TPU and a Confidential TPU