Connect with us

NEWS

Google Fields Four TPU Systems From Existing Chips

Google’s TPU 8tx and 8i Max are new systems, not new chips, and the confidential model runs on AMD EPYC hosts.

Published

on

Google added TPU 8tx and TPU 8i Max on November 9, 2026, the third and fourth systems in its eighth-generation TPU line. Neither unit is a new piece of silicon.

Amin Vahdat, Google’s senior vice president and chief technologist for AI and infrastructure, said the company rebuilt packaging, networking, and security around chips it already makes, after regulated buyers and long-context agent jobs asked for setups a single training or serving tray does not provide.

Google Calls Them Systems, Not Chips

Vahdat drew a hard line that most launch notes skip. A new chip, he wrote, means laying out billions of transistors on a wafer through a multi-year manufacturing cycle. A system is how Google groups compute, networking, storage, and security around that wafer.

That approach allows us to deliver specialized supercomputers in a matter of months rather than years.

Amin Vahdat, SVP and Chief Technologist for AI and Infrastructure, Google blog

TPU 8tx uses the same die as TPU 8t, the training part announced in April. TPU 8i Max puts two TPU 8i inference dies in one package. Google now counts four TPU systems in a single year, which Vahdat said the company had never done before.

That cadence used to run closer to one new processor every two years. The November drop is the proof of the modular bet: when software and customer jobs move faster than a tape-out, Google ships another tray instead of another die.

The Eighth Generation Now Has Four Jobs

At Google Cloud Next on April 22, 2026, the company split the generation on purpose. TPU 8t is the training tray. TPU 8i is the low-latency reasoning tray. The two November systems sit on those dies and take two jobs the April pair left open: secrets that cannot sit in cleartext RAM, and context windows that stall when memory bandwidth runs out.

Google’s own April figures still describe the silicon underneath. A TPU 8t superpod scales to 9,600 chips with two petabytes of shared high-bandwidth memory and 121 ExaFlops of compute. Each 8t chip carries 216 GB of HBM, 128 MB of on-chip SRAM, 12.6 PFLOPS at FP4, and 6,528 GB/s of HBM bandwidth. Each 8i chip carries 288 GB of high-bandwidth memory, 384 MB of SRAM, 10.1 PFLOPS at FP4, and 8,601 GB/s of HBM bandwidth.

THE FOUR EIGHTH-GENERATION TPU SYSTEMS

System Job Silicon and host Standout spec
TPU 8t Scale-up training TPU 8t die, Axion (Arm) 9,600-chip superpod, 216 GB HBM, 12.6 PFLOPS FP4
TPU 8i Low-latency reasoning TPU 8i die, Axion (Arm) 288 GB HBM, 384 MB SRAM, 8,601 GB/s
TPU 8tx Confidential inference, small-scale training TPU 8t die, 5th Gen AMD EPYC Up to 2x lower latency than TPU 8t
TPU 8i Max Long-context inference Two TPU 8i dies in one package 384 GB HBM, 768 MB SRAM, double 8i’s HBM bandwidth

Google said TPU 8t and TPU 8i offer up to 2.7 times better training performance and 80% better inference price-performance than the previous TPU generation. Those claims sit on the April dies. The November trays do not add a third architecture.

AMD Takes the Host Socket Axion Just Won

The confidential tray is also a host-CPU reversal. In April, Google said TPU 8t and TPU 8i were the first generation to run on its own Axion Arm-based CPU hosts, with more physical hosts per server and a NUMA layout meant to keep accelerators fed. TPU 8tx leaves that Arm board behind.

Google paired the 8t die with 5th Gen AMD EPYC processors and borrowed the Boardfly network from TPU 8i. AMD said those EPYC parts are the first cloud TPU with hardware-enforced confidential computing, and that they already host accelerators at more than 10 hyperscalers and frontier labs.

Lisa Su, AMD’s chair and chief executive, wrote that the company was proud to partner, “with AMD EPYC serving as the host processor for TPU 8tx.” Dan McNamara, AMD’s senior vice president for compute and enterprise AI, said the most capable AI systems need a CPU foundation that can match the accelerators they host.

Mark Lohmeyer, Google Cloud’s vice president for AI and computing infrastructure, put the buyer problem in one line: teams moving agents into production need to know data and intellectual property stay private, even from the machines underneath.

By pairing 5th Gen AMD EPYC processors with our 8th-generation TPU architecture inside the AI Hypercomputer, we are bringing hardware-enforced confidential computing to cloud TPUs for the first time.

Mark Lohmeyer, VP and GM for AI and Computing Infrastructure, Google Cloud, AMD newsroom

Axion still sits under the training and serving trays. The SKU that banks and agencies are supposed to buy runs on x86, because that is where AMD’s Secure Encrypted Virtualization already lives in Google Cloud’s confidential VM line.

What TPU 8tx Encrypts While Models Run

Data is usually encrypted on disk and on the wire, then unlocked in memory while a processor works on it. That gap is the whole product. TPU 8tx is built for regulated enterprises and governments that want to run frontier or proprietary models, and agents, next to their most sensitive records.

AMD said the host uses SEV, then extends that protection from the CPU to the attached TPUs. An accelerator’s identity and security setup are checked before it gets direct access to protected memory, so prompts, model weights, and intermediate data stay encrypted in CPU memory and in transit between host and chip, with hardware isolation while the TPU runs.

Google Cloud already sells hardware-based memory encryption in a trusted execution environment on confidential VMs, including AMD SEV and SEV-SNP. TPU 8tx is that idea moved onto the accelerator path. AMD also noted that Infinity Guard features vary by EPYC series and must be turned on by the cloud provider.

WHAT THE CONFIDENTIAL TRAY IS FOR

  • In-use isolation: Data and model weights stay protected and isolated while the chips are working, not only at rest or on the network.
  • One-cluster fine-tuning: Teams can fine-tune, distill, and run reinforcement learning on the same system without moving private data between clusters.
  • Latency claim: Google said Boardfly on 8t silicon delivers up to 2x lower latency than TPU 8t with a near-zero performance penalty for the security path.

Vahdat’s post does not name an attestation method, a threat model, or an independent audit a regulated buyer could run against that isolation promise. Until those artifacts exist, the security story is a vendor claim sitting on AMD’s SEV design and Google’s packaging.

Two 8i Chips Share One Package

TPU 8i Max is aimed at a different stall. Vahdat called it the memory wall: when dozens of agents work across a whole codebase, or hours of video, memory bandwidth gives out before math does. Google combined two low-latency TPU 8i chips on a single package and said that setup delivers double the compute, double the SRAM, and double the HBM bandwidth of TPU 8i, with twice the silicon in the existing rack footprint.

Bandwidth is the part that actually doubles. SRAM moves from 384 MB to 768 MB. HBM bandwidth doubles off 8i’s 8,601 GB/s. HBM capacity does not double. The packaged figure is 384 GB, one-third more than a single 8i’s 288 GB, which is the honest reading of a two-die tray that still has to fit in the same rack slot.

Google told buyers that jobs which must stay inside the confidential domain should stay on TPU 8tx for inference and serving. Max is the density and context tray, not the secrets tray. TPU 8i already cut Boardfly’s maximum network diameter by more than 50% versus the training torus, doubled interconnect to 19.2 Tb/s, and added a Collectives Acceleration Engine that Google said reduces on-chip latency by up to 5x. Max inherits that serving fabric and feeds it more local bandwidth.

A Customer List That Reaches On-Prem Racks

The SKU split follows who now pays. Google said it first built TPUs for internal jobs such as speech and translation. Eight generations later the same silicon trains and serves Gemini, runs agent workloads, sells as cloud capacity, and, Google said, now ships so customers can put TPUs in their own data centers.

Vahdat wrote that TPUs are a rapidly growing business with nearly 100 customers, including frontier labs, trading firms, retailers, and universities. Citadel Securities was named in the April launch as a TPU user. Feedback, he said, now comes from those outside groups as well as Google DeepMind, and the two new systems exist because businesses asked for accelerators that can reason across huge context windows and because some highly regulated customers asked for encryption built for their jobs.

That is a different company from the one that used to tape out a single TPU and run it in-house. A four-system generation only pays off if those nearly 100 accounts, and the on-prem racks behind them, will buy a security tray and a context tray instead of renting a generic GPU and wrapping it in software controls.

PyTorch Runs Native, and Idle Agents Get Parked

Hardware is the smaller half of the November note. Google spent as much of the post on software that is supposed to make a TPU feel like a general accelerator, which is the barrier that still keeps most PyTorch shops on GPUs.

THE SOFTWARE GOOGLE TIED TO THE NEW TRAYS

  • TorchTPU: Google said it open-sourced TorchTPU so PyTorch code and Hugging Face models run natively on TPUs, with TPU backends for vLLM and SGLang.
  • TPU Sync: The open-source model-state layer is claimed to cut KV-cache transfer latency and time to first token by more than 90% by tiering caches off chip.
  • Agent Compute Engine: Because agents pause for tools, databases, or a human, the software suspends idle sessions so parked work does not keep burning capacity.

Cloud Pathways is now generally available on Google Kubernetes Engine, which Google said lets customers drive large TPU slices from a single controller. GKE dynamic slicing, also generally available, uses optical circuit switching to build or split topologies on the fly, with Google claiming job startup up to 5x faster and fault recovery up to 4.5x faster. TPU 8t and TPU 8i were already rated at up to 2x better performance-per-watt than Ironwood, with TPU 8t aimed at over 97% goodput.

Google did not publish a price, a region list, or a general-availability date for TPU 8tx or TPU 8i Max. The dies underneath were announced in April. The new trays are configurations of those dies, and the confidential one needs AMD’s host stack plus an attestation story Google has not yet put on the page. Regulated buyers will wait for all three before they move a workload that cannot leak.

Harry is the editor of THE iBULLETIN, an independent publication he owns and runs. He has been in journalism for ten years, first reporting and later editing, and much of what the site covers now begins in its inbox. Reader mail is read in full, every message of it. A tip is treated as a lead to be verified, not a story to be printed, and a challenge to a published fact is checked against the original filing, statement or transcript within the day, with the article corrected under a public policy if the reader is right. Questions that several readers ask become articles. That exchange feeds coverage of news, business and technology, of science and sports, and of entertainment, lifestyle, travel, auto and gaming, written for readers spread across many countries rather than one. Harry works from primary sources and checks each number himself before publication, and he would rather run a shorter story than an unconfirmed one. The address for all of it, tips, corrections and questions alike, is support@theibulletin.com.

Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Trending