Key points
- Google adds two eighth-generation TPU systems
- Why Google is pairing TPUs with AMD chips
- What Google's TPU 8i Max is built for
Google is adding two versions to its eighth-generation lineup of tensor processing units, or TPUs, bringing it to four systems, Bloomberg News reported on November 9. The TPU 8tx is designed for confidential computing and works alongside chips from Advanced Micro Devices (AMD). The TPU 8i Max is built for AI tasks that need long context windows.
Both systems are due next year, according to Amin Vahdat, Google's chief technologist for AI and infrastructure.
Alphabet (GOOGL) shares were up 3% at $293.26 in premarket trading as of 8:11 a.m. ET. They were up about 1.8% just before the report came out at 7:30 a.m.
The TPU 8tx is aimed at government agencies and regulated industries
Confidential computing is designed to protect data while it's in use, including by isolating workloads and protecting their memory. Google is aiming the TPU 8tx at government agencies and companies in regulated industries, and the system works alongside AMD chips as part of a partnership between the two companies, according to Bloomberg's report.
Google and AMD have worked together on confidential computing before. Google Cloud offers confidential virtual machines that run on AMD's EPYC server processors. Those machines use an AMD feature that keeps each virtual machine's memory encrypted, according to a 2022 AMD release.
Google said the TPU 8t and TPU 8i it announced in April use its own Axion Arm-based CPUs.
The TPU 8i Max is built for long context windows
A context window is the amount of information an AI model can take in at once when it answers a request. Bloomberg reported that the new chips will help models "access a wider range of information when handling queries."
The 8i Max combines two TPU chips in one package for more computing power and memory capacity. The standard TPU 8i, Google's chip for running AI models, pairs 288 GB of high-bandwidth memory with 384 MB of on-chip SRAM, according to Google's April announcement.
Google's eighth-generation TPU lineup now has four systems
| System | Built for | Expected availability |
|---|---|---|
| TPU 8t | Training AI models | Later in 2026 |
| TPU 8i | Running AI models (inference) | Later in 2026 |
| TPU 8tx | Confidential computing, with AMD chips | 2027 |
| TPU 8i Max | Long context windows, two chips per package | 2027 |
Google said in April that the TPU 8t and TPU 8i would be generally available later in 2026, and Bloomberg reported that the two new systems are due next year. Vahdat said each of the four systems is tailored to different needs, according to Bloomberg.
Google began offering TPUs in outside data centers in the second quarter. Oracle (ORCL) plans to offer Google's TPUs on its cloud starting in 2027.
