Chips & Compute

Google ships four TPU systems in one year with 8i Max and 8tx

Google adds TPU 8i Max and TPU 8tx, its fourth and fifth launches this year, targeting confidential computing and large-context agent workloads alongside a expanded software stack.

Google expands TPU line-up with two new AI systems
Google expands TPU line-up with two new AI systemsgeralt / pixabay
By Sophie Lindqvist4 min read

Why it matters

  • Google launched four TPU systems in a single year, capped by TPU 8i Max and TPU 8tx.
  • TPU 8tx delivers up to half the latency of TPU 8t with near-zero performance penalty.
  • TPU 8i Max doubles compute, SRAM and HBM bandwidth versus TPU 8i by pairing two chips per package.
  • TPU Sync cuts key-value transfer latency and time to first token by more than 90%.
  • Nearly 100 customers now use TPUs, including frontier labs, trading firms, retailers and universities.

Google has released two new eighth-generation Tensor Processing Units, TPU 8i Max and TPU 8tx, bringing its TPU count to four systems launched this year. TPU 8tx handles secure, confidential inference; TPU 8i Max targets workloads that need large context windows and more compute packed into the same rack space.

Google says nearly 100 customers now run on TPUs, a list that includes frontier laboratories, trading firms, retailers and universities. The company frames the launches as a response to demand across model training, inference and autonomous agents. The stakes are straightforward: cloud providers and chip suppliers are racing to supply AI infrastructure that offers lower latency, bigger memory footprints and tighter data controls, and Google is pushing specialised silicon rather than one-size-fits-all chips.

What does TPU 8tx change for regulated customers?

TPU 8tx runs on the same silicon as TPU 8t, paired with fifth-generation AMD EPYC CPU hosts. Google pitches it at regulated enterprises and government users that want to run models and agents next to sensitive data.

The key mechanism is hardware-level confidential computing. According to Google, TPU 8tx protects data and model weights while they are actively in use, not just at rest. It also reuses the Boardfly network introduced for lower-latency reasoning workloads, and Google claims it delivers up to half the latency of TPU 8t with near-zero performance penalty.

The practical effect: customers can run inference and smaller-scale training on one system. Google says organisations can fine-tune models with private data, distil larger models into smaller ones, and train agents through reinforcement learning without shuttling sensitive data between separate clusters.

Why did Google double up chips in TPU 8i Max?

TPU 8i Max takes aim at what Google calls the memory bottleneck in AI systems. The design combines two low-latency TPU 8i chips into a single package, which doubles compute, SRAM and high-bandwidth memory bandwidth compared with a single TPU 8i, according to the company.

Google says the result suits workloads where multiple agents operate across large code bases or extended video inputs. The packaging matters as much as the silicon: twice the silicon fits into the existing rack footprint, so data centres get more compute per square foot and per watt.

How does four-chip product line fit together?

The eighth-generation TPU family now covers four distinct jobs:

  • TPU 8t — large-scale training
  • TPU 8i — lower-latency reasoning
  • TPU 8tx — confidential computing
  • TPU 8i Max — high-context workloads

Google describes this as a system-level approach. Instead of waiting for a full new chip generation, it modifies the surrounding compute, networking, storage and packaging around existing silicon to create more specialised systems. That lets the company respond faster to changing customer requirements, it said.

"With computing moving from human to agent scale, businesses are asking for accelerators that can reason across AI models' massive context windows," Google said. "In addition, some highly regulated customers are asking for specialized encryption capabilities for unique workloads."

What's new on the software side?

Hardware is only half the announcement. Google also detailed a broader push to make TPUs easier to use for developers and cloud customers.

PyTorch code and Hugging Face models now run natively on TPUs through the open-source TorchTPU project. TPU back ends are available for open-source inference engines including vLLM and SGLang.

A TPU agent suite, currently in preview, plugs into coding tools to automate onboarding, diagnose bottlenecks and tune kernels. For production use, Google's open-source XLA Serving Engine lets customers train models in Python and serve them in a C++ runtime across TPUs, GPUs and CPUs.

Two other updates target cost and speed:

  • Agent Compute Engine suspends and resumes idle agent sessions to cut infrastructure costs.
  • TPU Sync manages model state across high-bandwidth memory, host memory, storage and network resources. Google says it can reduce key-value transfer latency and time to first token by more than 90%.

Cloud Pathways is now generally available on Google Kubernetes Engine, extending distributed runtime tools Google used internally for Gemini to Google Cloud customers. Separately, GKE dynamic slicing lets users reconfigure TPU topologies on demand, which Google says cuts job start-up times by up to five times and improves fault recovery by up to 4.5 times.

Where does AI meet chip design?

Google also disclosed that it uses AI tools from Google DeepMind inside its own chip design process. That includes AlphaChip reinforcement learning systems and Gemini-based design tools, which the company says have cut parts of chip layout work from weeks to hours.

"Shipping four TPU systems and a unified, open software stack in a single year only happens when silicon architects, networking engineers, compiler teams and Google DeepMind researchers build side by side," Google said.

The cadence is the real signal here. By varying packaging, networking and host systems around shared silicon, Google can segment the AI accelerator market into specialised niches faster than a traditional chip-generation cycle allows, and the near-100-customer base suggests enterprises are buying that approach.

Source: itbrief.com.au

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.

749 articles

Related articles

  1. Google Adds TPU 8Tx and TPU 8I Max to Eighth-Generation Lineup
  2. Google’s new TPUs target secure, high-memory AI agents
  3. Google Ships TPU 8tx With Encrypted In-Use AI Compute and 8i Max
  4. Google ships four TPU systems in one year to challenge Nvidia
  5. AMD 5th Gen EPYC processors host Google Cloud's confidential TPU 8tx

« Previous articleNext article »