AMD Epyc to power Google Cloud TPU 8tx with hardware confidential computing - Chosunbiz
Google Cloud has launched its TPU 8tx AI accelerator system, which pairs eighth-generation TPU architecture with fifth-generation AMD Epyc processors to enable hardware-based confidential computing.
The TPU 8tx marks the first time Google Cloud's tensor processing units support hardware-based confidential computing. AMD's fifth-generation server CPU handles data preparation, workload distribution, and communication with the TPU accelerator, while the TPU performs the AI computation.
Confidential computing technology protects data not only when stored or transmitted but also while being processed. This system applies AMD's Secure Encrypted Virtualization (SEV) technology to encrypt and protect AI models, data, and information moving between the CPU and TPU.
Enhanced Security Features
Before the accelerator accesses protected memory, the system verifies the device's identity and security configuration. This protects user input prompts, AI model weights, and intermediate computation data while applying hardware-based isolation during job execution.
When bringing AI agents into real production environments, you need confidence that confidentiality is maintained so the underlying infrastructure cannot access data or intellectual property.
Strategic Partnership
AMD and Google Cloud have collaborated across multiple generations of Epyc CPUs. The processors are used in Google Cloud's C4D, N4D, H4D, Z4D, and G4 virtual machines, with Google security experts reviewing firmware for AMD security processors.
Epyc processors have been adopted as host CPUs supporting GPUs and custom AI accelerators at more than 10 major global cloud providers and leading AI labs, including systems with AMD Instinct GPUs, Nvidia GPUs, and ASICs.
Related
