{
  "id": 19643395,
  "title": "Google reworks TPU silicon into specialized systems for agents and security",
  "url": "https://urgent.news/2026/11/12/google-reworks-tpu-silicon-into-specialized-systems-for-agents-and",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-11-12T11:00:00.000Z",
  "source": {
    "name": "Tom's Hardware",
    "slug": "tom-s-hardware",
    "url": "https://www.tomshardware.com/tech-industry/artificial-intelligence/google-adds-two-new-specialized-chips-for-ai-agents-and-security-to-8th-gen-tpu-lineup-tpu-8i-max-puts-two-chips-on-one-package-for-long-context-work-tpu-8tx-swaps-arm-hosts-for-amd-epyc"
  },
  "original_language": "en",
  "account": "Google has unveiled its latest advancements in Tensor Processing Unit (TPU) silicon, introducing specialized systems tailored for agents and security purposes. The company's eighth-generation TPU family now includes TPU 8tx and TPU 8i Max, designed to cater to confidential, hardware-isolated inference workloads for regulated enterprises and governments, as well as long-context agentic workloads.\n\nThe TPU 8tx replaces Arm-based Axion hosts with a 5th Gen AMD EPYC CPU, adopting a Boardfly network, while the TPU 8i Max combines two 8i chips on one package, doubling its compute, SRAM, and HBM bandwidth compared to the original 8i. This modular approach allows Google to deliver specialized systems more quickly, bridging the compute gap with competitors like Nvidia and AMD.\n\nTPU 8i Max offers double the theoretical compute capacity and more on-chip SRAM and HBM bandwidth, achieving approximately 20.2 FP4 PFLOPS, 768 MB SRAM, and about 17.2 TB/s HBM bandwidth. In comparison, Nvidia's Rubin R200 GPU boasts 50 FP4 PFLOPS for inference and 22 TB/s of HBM4 per socket, while AMD's MI455X delivers 40 FP4 PFLOPS.\n\nThe TPU 8tx is specifically engineered for secure and confidential inference, along with small-scale training. Its hardware-based isolation ensures data protection, making it suitable for regulated enterprises and governments working with sensitive models. Google has also developed software tools, such as TPU Sync and Agent Compute Engine, to optimize performance and memory usage for these workloads.",
  "summary": "Google’s TPU 8i Max and TPU 8tx reuse TPU 8i and TPU 8t silicon to target long-context agents and confidential inference.",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}