# Google Rejiggers TPU Accelerators For Broader AI Inference Work

DevFeed: [Google Rejiggers TPU Accelerators For Broader AI Inference Work](<https://devfeed.tech/articles/google-rejiggers-tpu-accelerators-for-broader-ai-inference-work-124816.md>)

Original publisher: [Read original article](<https://www.nextplatform.com/compute/2026/11/09/google-rejiggers-tpu-accelerators-for-broader-ai-inference-work/5308534>)

Author: Timothy Prickett Morgan

Published: 2026-11-09T18:06:49Z

Content type: news

Language: en

Sources: [The Next Platform: In-depth coverage of high end computing](<https://devfeed.tech/sources/the-next-platform-in-depth-coverage-of-high-end-computing.md>)

Topics: [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [AI Factory](<https://devfeed.tech/topics/ai-factory.md>), [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [Google Cloud Platform (GCP)](<https://devfeed.tech/topics/google-cloud.md>), [Google](<https://devfeed.tech/topics/google.md>)

Tags: [accelerators](<https://devfeed.tech/tags/accelerators.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-inference](<https://devfeed.tech/tags/ai-inference.md>), [broadcom](<https://devfeed.tech/tags/broadcom.md>), [compute](<https://devfeed.tech/tags/compute.md>), [google](<https://devfeed.tech/tags/google.md>), [inference](<https://devfeed.tech/tags/inference.md>), [mediatek](<https://devfeed.tech/tags/mediatek.md>), [sunfish](<https://devfeed.tech/tags/sunfish.md>), [tpu](<https://devfeed.tech/tags/tpu.md>), [tpu-8i](<https://devfeed.tech/tags/tpu-8i.md>), [tpu-8t](<https://devfeed.tech/tags/tpu-8t.md>), [zebrafish](<https://devfeed.tech/tags/zebrafish.md>)

## AI overview

Google's new TPU variants target distinct workloads through changes to interconnect, memory, and host processors. The TPU 8tx adapts the training-focused TPU 8t for inference: its boardfly fabric reduces pod size from 9,600 to 1,152 accelerators and is described as having about half the average latency of the 3D torus, while an AMD Epyc replaces Google's Axion CPU to provide hardware root of trust and encryption acceleration. The TPU 8i Max pairs two Zebrafish chips to provide more shared memory for long-context inference, though memory capacity rises only 33% per socket and memory per unit of compute falls. Google's TPU Sync software moves an agent's working memory off the TPU while paused and restores it on resume, freeing accelerators for other tasks; the excerpt says it also supports TPU 5 through TPU 8, but gives no throughput measurements. The modified systems are expected in 2027. Their cloud pricing and the scale and performance of TPU Sync remain uncertain.

## Source excerpt

Sometimes you need a new chip, and sometimes you need a new way to configure the chips you have to b ...