blog.google via Hacker News

Google étend sa 8ᵉ génération de TPU avec 8tx et 8i Max

5 médias qui couvrent ce sujet

TL;DR

  • All four TPU 8th-gen variants share one die; Google achieves specialization through packaging, host CPU pairing, and security enclave configuration.
  • TPU 8i Max stacks two chips per socket, doubling HBM bandwidth and SRAM for 1M-token context windows without increasing rack footprint.
  • AMD EPYC Turin pairing is a notable hyperscaler design win; EPYC rarely appears in Google's custom accelerator stack.

Google greffe deux systèmes supplémentaires sur sa huitième génération de TPU. Dans un billet signé Amin Vahdat, SVP and Chief Technologist, AI and Infrastructure, la société présente le TPU 8tx, « optimized for secure and confidential inference », et le TPU 8i Max, taillé pour les très longs contextes agent. « This is the first time we have introduced four TPU systems in a single year », écrit Vahdat.

Le 8tx reprend le silicium du 8t : « Built on the same silicon as TPU 8t and paired with 5th Generation AMD EPYC », indique le billet, la protection confidentielle venant côté hôte. Google avance « up to 2x lower latency compared to TPU 8t with a near-zero performance penalty ». Le 8i Max, lui, soude deux puces 8i sur un même boîtier pour offrir « double the compute, double the SRAM and double the HBM bandwidth of TPU 8i ». Google motive ce format par les charges d'agents : « When dozens of AI agents work together across an entire software codebase or hours of video ».

Le billet rappelle aussi que Google a « open-sourced TorchTPU » le mois précédent, de sorte que « PyTorch code and Hugging Face models now run natively on TPUs ». Les deux premiers systèmes 8ᵉ gén, 8t et 8i, servent de base de comparaison avec « up to 2.7 times better training performance and 80% better inference price performance » face à la génération antérieure.

L'annonce s'inscrit dans un flux dense d'actualités infrastructures IA que nous suivons (264 alertes en 90 jours), et prolonge l'engagement de 17 Md$ de Google au Wyoming publié quelques jours plus tôt.

Ce qu'en disent les autres médias

Couverture consolidée 24h après publication

  1. SiliconANGLE Lire →

    Contextualizes memory bottlenecks as the primary design driver for 8i Max, linking the 1M-token context requirement directly to the dual-chip packaging decision.

    TPU 8i Max fits twice the silicon in the same rack footprint, lowering the floor space it takes per square foot and per watt in data centers.
  2. ServeTheHome Lire →

    Frames AMD EPYC Turin pairing as a hyperscaler design win and details Boardfly network topology as the switching fabric enabling the 8tx's security architecture.

    The 8tx keeps that active data protected through confidential computing, which encrypts data in use inside an isolated trusted execution environment.
  3. Interesting Engineering Lire →

    Leads with the 80% inference price-performance gain and covers TorchTPU, vLLM backends, and TPU Sync as the software layer needed to convert hardware specs into developer adoption.

    TPU 8tx uses the same silicon as TPU 8t, paired with fifth-generation AMD EPYC CPU hosts.
  4. MLQ.ai Lire →

    Tracks the full 2026 TPU release cadence and notes NVIDIA already offers comparable confidential compute, sharpening the competitive question to execution speed and software ecosystem.

    TPUs now serve nearly 100 customers across frontier labs, trading firms, retailers and universities.