Science & Tech · Big Tech & Internet · 1 hr ago
Google's confidential AI system uses AMD Epyc CPUs instead of Axion
Google introduced the TPU 8tx, a system for running AI work with confidential computing protections.
It pairs Google's TPU 8t accelerator with fifth-generation AMD Epyc processors, rather than Google's own Axion CPU.
The system is aimed at governments and regulated businesses that want to work with private data without moving it between clusters.
AMD says its security technology helps isolate the system and protect prompts, model weights and other data while the TPU is running.
Google has not explained the choice, but its current confidential-computing services support AMD and Intel machines, not Axion, suggesting its audited security setup is limited to x86 processors.
Google says the TPU 8tx can deliver up to twice the performance of the TPU 8t on latency, with nearly no performance penalty, while adding hardware-enforced protection for AI workloads.
Google’s TPU 8tx pairs its AI accelerator with fifth-generation AMD Epyc “Turin” CPUs for confidential AI inference.
It is the first cloud TPU system described by Google and AMD as offering hardware-enforced confidential computing.
The system uses Google’s TPU 8t silicon and adds a lower-latency network used by the TPU 8i.
AMD’s Secure Encrypted Virtualization works with the TPU to protect data and extend the protected environment to the accelerators.
Google has not directly explained why it used AMD CPUs instead of its own Axion processors.
- Who
- Google and AMD. Google’s TPU 8tx uses AMD Epyc CPUs.
- What
- Google introduced a TPU system for confidential AI inference with hardware-enforced confidential computing.
- When
- The article was published on November 15, 2026. Google launched its eighth-generation TPU hardware in April 2026.
- Where
- Google Cloud.
- Why
- Google has not directly explained its choice. The article says its confidential computing stack may be limited to audited x86 systems.
This story does not have two clearly opposing sides.
Google launched its eighth-generation TPU hardware, including the TPU 8t for training and TPU 8i for inference.
TechRadar reported that Google’s TPU 8tx pairs TPU 8t silicon with AMD Epyc CPUs for confidential AI inference.
- System
- TPU 8tx
- Host processor
- 5th Gen AMD Epyc “Turin
- Accelerator silicon
- TPU 8t
- Memory
- 216GB of HBM
- Compute
- 12.6 petaflops of FP4 compute
- Security technology
- AMD Secure Encrypted Virtualization (SEV)
What people are saying
keeping sensitive data and model weights protected and isolated even while in active use.