Skip to the stories
Windows & PowerShell

Google's TPU 8tx and TPU 8i Max Split the Eighth-Gen Lineup by Workload

  • Brian Beeler
  • 599 views

Google adds TPU 8tx for confidential computing and TPU 8i Max for dual-chip long-context inference, with no cloud availability dates or pricing yet.

Google's TPU 8tx and TPU 8i Max Split the Eighth-Gen Lineup by Workload

Google has filled out its eighth-generation TPU lineup with two new systems, and the more useful read for infrastructure teams is that each one targets a specific constraint rather than a raw speed number. TPU 8tx is a confidential computing system built on TPU 8t silicon. TPU 8i Max places two TPU 8i chips on a single package for long-context inference. Both join the TPU 8t training chip and TPU 8i inference chip that Google introduced at Cloud Next in April. Google says this is the first year it has introduced four TPU systems.

Google TPU 8tx and TPU 8i Max boards rendered side by side, with each TPU 8i Max board carrying two chips per package

Google describes both additions as new systems built on existing silicon. The company changed the packaging, high-speed memory, networking, and security around the chips to suit each workload. Google credits that modular approach with getting specialized configurations out in months rather than a full chip cycle. According to Google, TPUs now have nearly 100 customers, including frontier labs, trading firms, retailers, and universities. Some of those customers can now deploy TPUs directly in their own data centers.

TPU 8tx: Confidential Computing on the TPU 8t Die

TPU 8tx uses the same silicon as TPU 8t and pairs it with 5th Generation AMD EPYC host CPUs. Confidential computing is built into the hardware, so data and model weights stay protected and isolated while they're in use. That's the gap this system is meant to close. Plenty of teams can encrypt data at rest and in transit, but keeping a proprietary model's weights shielded during computation has been harder. Google lists PCIe link encryption, cryptographic attestation, and secure memory enclaves among the security features. It's aiming the system at regulated enterprises and public-sector organizations that run frontier or proprietary models and agents alongside sensitive data.

Google TPU 8tx board inside a secure enclosure graphic listing PCIe link encryption, cryptographic attestation, and secure memory enclaves

TPU 8tx also adopts the low-latency Boardfly network Google built for TPU 8i. Google says this delivers up to 2x lower latency than TPU 8t with a near-zero performance penalty. The pairing is what makes the system practical for secure inference and small-scale training. A team could fine-tune on private data, distill a large model into a smaller one, or train agents with reinforcement learning without moving sensitive datasets between separate clusters. For many organizations, that data-movement problem is the real cost, more than the compute itself.

TPU 8i Max: Two Inference Chips in One Package

TPU 8i Max is aimed at memory bandwidth. Google points to workloads such as dozens of agents working across an entire codebase, or hours of video held in context at once. The design puts two TPU 8i chips on one package, which Google says doubles compute, on-chip SRAM, and HBM bandwidth over TPU 8i. Google listed 384MB of SRAM for TPU 8i at launch, so the doubled figure works out to 768MB per 8i Max package. Google hasn't stated the package's HBM capacity, so that's a number to request directly before you size anything.

The rack argument is the part that matters for facilities planning. Because 8i Max fits twice the silicon into the existing rack footprint, Google says it delivers more compute per square foot and per watt. If that holds in your environment, the upgrade path looks less like a new hall and more like a denser fit-out. I'd treat the density claim as Google's own until you can measure power and cooling against your actual load profile.

Google TPU 8i Max render showing four boards, each with two TPU 8i chips on a single package

Google now maps the four eighth-generation systems to stages of agent work. TPU 8t handles scale-up training. TPU 8i handles low-latency reasoning. TPU 8tx handles confidential computing. TPU 8i Max handles massive context windows and the highest compute density per rack. That's a clean way to think about procurement: match the system to the bottleneck, not to the generation number. Google hasn't said when TPU 8tx or TPU 8i Max will reach Google Cloud customers, and it hasn't published pricing for either system. Until it does, capacity planning is guesswork.

Google's eighth-generation TPU lineup: TPU 8t for scale-up training, TPU 8i for low-latency reasoning, TPU 8tx for confidential computing, and TPU 8i Max for massive context and density

TPU Sync, Cloud Pathways, and GKE Dynamic Slicing

The software announcements are arguably more immediate for existing TPU users. Google open-sourced TPU Sync, a C++ library that treats chip HBM, host DRAM, shared storage, and the data center network as one memory hierarchy. It tiers KV caches off the chip, streams context from prefill nodes to decode nodes, and synchronizes reinforcement learning weights from trainers to samplers. Google says TPU Sync cuts KV transfer latency by more than 90 percent and reduces time to first token by an average of 90 percent on open-weight mixture-of-experts models ranging from 235B to 480B parameters. Those are Google's test results on those model sizes, and they're worth reproducing on your own workloads before you rely on them.

Cloud Pathways, the Google DeepMind runtime behind Gemini, is now generally available on Google Kubernetes Engine. It uses a single controller to orchestrate TPU systems across multiple slices for pretraining, reinforcement learning, and multihost inference. It keeps long-running jobs going through hardware disruptions, using sub-minute slice hot-swapping and elastic resharding. For anyone who has lost a multi-day training run to a failed host, that recovery behavior is the feature to look at first.

GKE dynamic slicing is also generally available. It uses optical circuit switching to assemble TPU cubes into larger topologies, or split them into smaller isolated workloads, without recreating node pools. Google says this cuts job startup times by up to 5x and speeds fault recovery by up to 4.5x. A TPU agent suite is in preview and plugs into coding tools such as Claude Code, Cursor, and Google Antigravity to onboard workloads, diagnose bottlenecks, and autotune kernels. Google also released a new open-source XLA Serving Engine, which serves Python-trained models from a C++ runtime across TPUs, GPUs, and CPUs.

Here's the thing: the hardware news is the headline, but the TPU 8tx confidential option and the GKE slicing changes are what will shape near-term decisions. If your regulated workloads have been kept off shared accelerators, the attestation and enclave features are worth tracking closely. The open questions are cloud availability dates, pricing, and HBM capacity for 8i Max. Watch for those before the next change window.

Google Cloud TPU

Source: Brian Beeler · www.storagereview.com

Neighbouring cells