TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Google has added the TPU 8tx, designed for confidential computing, and the TPU 8i Max, aimed at memory-intensive workloads, to its eighth-generation accelerator lineup. The company says it has shipped four TPU systems in one year, expanding a platform it is offering to external customers as competition over AI inference hardware grows.
Google has added two systems to its eighth-generation TPU lineup—the TPU 8tx for confidential computing and the TPU 8i Max for memory-intensive workloads—bringing the company’s reported total to four TPU systems shipped in one year. The expansion gives Google more configurations to offer businesses and public-sector organizations as it seeks to make its in-house AI accelerators a platform for customers beyond Google.
The new systems use existing TPU silicon, rather than introducing a new chip generation. The changes are in system design, including packaging, networking and memory configurations. Google presents this modular approach as a way to introduce hardware variants faster than developing a new processor generation, though the source report does not provide independent performance tests or a detailed release calendar.
The TPU 8tx is intended for organizations handling sensitive workloads. Google says it keeps data and model weights protected and isolated while they are being processed, addressing the exposure that can occur when encrypted information must be decrypted for computation. It uses the Boardfly network, which Google developed for real-time inference. The company says it can reduce latency by up to twice as much as the TPU 8t; the precise test conditions and comparison details are not supplied in the source material.
The TPU 8i Max is designed for workloads that require more memory capacity and bandwidth, including AI agents working across large codebases or lengthy video. Google says it places two TPU 8i chips on one package, providing double the compute power, memory and memory bandwidth in the same rack format. Separately, Google has released TorchTPU as open source and says PyTorch code can run on TPUs with minimal changes. The source also reports support for inference frameworks vLLM and SGLang.
Google Broadens Its TPU Customer Offer
The additions signal that Google is treating TPUs as more than infrastructure for its own services. The report says the company now has nearly 100 external TPU customers, though it does not identify them or describe how much capacity they use. More configurations could help potential customers match a system to particular requirements, such as keeping sensitive data protected during processing or serving memory-heavy applications.
The shift also reflects the growing importance of AI inference—running trained models for everyday use—alongside the costly process of training them. Specialized accelerators may appeal to organizations seeking alternatives for specific workloads, but the source does not establish that the new TPUs are cheaper or faster than competing products in comparable deployments. Google’s claims will matter most to customers if they are borne out in independent testing and available capacity.
Google is competing in a market where Nvidia remains a major supplier of AI accelerators. The source report also points to chip efforts from Amazon, Microsoft and OpenAI, and says Nvidia has strengthened its inference push through a Groq-related acqui-hire and licensing arrangement. These developments show that major technology companies are pursuing more control over the hardware used to run AI services.
Products Worth Considering
Samsung Galaxy Tab A9+ 11-Inch Android Tablet, 64GB, Graphite
Logitech MX Ergo S Advanced Wireless Trackball Mouse USB-C – Graphite
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Raspberry
As an affiliate, we earn on qualifying purchases.
From Internal Chips to Cloud Customers
Google developed TPUs as its own application-specific accelerators and has increasingly positioned them for outside users. The move matters because customers choosing hardware for AI services may weigh performance, software compatibility, data-protection features and access to cloud capacity—not only the processor itself.
The four systems described in the report are variants within the eighth-generation family, not four separate chip generations. The latest additions extend the range of system configurations around existing silicon. Google’s open-source TorchTPU effort and support for frameworks including vLLM and SGLang are intended to make adoption easier, but the source provides no independent assessment of compatibility or the engineering work required for particular applications.
“The TPU 8tx can deliver up to twice the latency reduction compared with the TPU 8t.”
— Google, as reported by The Decoder
Performance and Availability Still Open
The source report does not specify when customers can access each new system, how much capacity Google plans to make available, or what the pricing will be. It also does not name the nearly 100 external TPU customers or say how many use the eighth-generation systems.
Google’s claims about latency and doubled resources are not accompanied in the material by independent benchmarks, detailed test conditions or comparisons with Nvidia systems. The report also does not explain the exact security architecture behind the TPU 8tx, how its protections are audited, or which regulatory and compliance requirements it supports. Those details will help customers judge whether the systems meet their operational needs.
Customer Access and Benchmarks
The next useful developments will be availability details, pricing and independent evaluations of the two systems. Customers will also need information about which Google Cloud regions and configurations will support them, and whether the claimed protections and performance hold across real workloads.
For Google’s broader strategy, evidence of customer adoption will be as important as the hardware announcement. The company’s reported external customer count indicates interest, but the source does not provide usage, revenue or deployment data. Comparisons based on public benchmarks and customer experience may clarify how far the expanded TPU portfolio can compete with Nvidia and other accelerators for inference workloads.
Key Questions
What are the two new Google TPU systems?
Google added the TPU 8tx, aimed at confidential computing, and the TPU 8i Max, designed for memory-intensive workloads. Both are system configurations using existing TPU silicon, according to the report.
Does this announcement introduce a new TPU chip generation?
No. The report says the systems use existing silicon and differ through packaging, networking and memory configurations. They expand the eighth-generation lineup rather than establish a new chip generation.
What does Google claim the TPU 8tx does?
Google says the TPU 8tx protects and isolates data and model weights while they are being processed, a feature intended for sensitive workloads. It also claims lower latency than the TPU 8t, but the source does not provide independent tests or full comparison conditions.
How does the TPU 8i Max differ from the TPU 8i?
Google says the 8i Max combines two TPU 8i chips on one package, doubling compute power, memory and memory bandwidth within the same rack format. The report does not include independent performance results.
When will customers be able to use the new systems?
The source material does not give a customer availability date, pricing or capacity plan. Those details remain to be announced or confirmed.
Source: the-decoder.com
