Skip to main content

Google Cloud TPU 8t (Trillium 2 / Training-Dedicated)

Overview

Google TPU 8t (codenamed Trillium 2) is the latest training-dedicated TPU, announced on 2026-04-22 (forming a split architecture with the simultaneously announced TPU 8i inference-dedicated TPU). It features 216GB HBM (12% more than TPU v7 Ironwood), 6,528 GB/s bandwidth, and an integrated Arm Axion CPU (Google's custom 64-core Arm processor).

TPU 8t is the core training chip for Google Gemini 3 / Gemini 4 frontier models, with key improvements over TPU v7 Ironwood in training paradigm optimization (MoE training, long-context training, RLHF post-training).

Core Specifications

ItemSpecification
ArchitectureTPU 8t (Trillium 2)
TypeTraining-dedicated (distinct from 8i inference-dedicated)
BF16 Compute (dense)~3,500 TFLOPS (estimated, ~50% higher than Ironwood's 2,307 TFLOPS)
FP8 Compute (dense)~7,000 TFLOPS
HBM Capacity216 GB
HBM Bandwidth6,528 GB/s
ICI Interconnect1,400 GB/s (bidirectional)
DCN Bandwidth200 Gbps (estimated)
Integrated CPUArm Axion (Google custom, 64-core)
Pod Size9,216 chips (estimated)
Topology3D Torus
Announcement2026-04-22

📌 8t naming: TPU 8th-gen + t = training. 8t and 8i are same-generation; 8t is for training only.

TPU 8t vs TPU v7 Ironwood (Training Comparison)

MetricTPU v7 IronwoodTPU 8tImprovement
TypeTraining + InferenceTraining-dedicatedType split
BF16 Compute2,307 TFLOPS~3,500 TFLOPS (estimated)1.5×
FP8 Compute4,614 TFLOPS~7,000 TFLOPS1.5×
HBM Capacity192 GB216 GB1.13×
HBM Bandwidth7,380 GB/s6,528 GB/sSlight decrease
ICI Interconnect1,200 GB/s1,400 GB/s1.17×
Integrated CPUNoneArm Axion 64-coreNew
Announcement2025-112026-04-22

💡 TPU 8t bandwidth slightly decreased (7,380 → 6,528 GB/s) but compute increased 50%, indicating Google traded some bandwidth for higher compute on 8t (better suited for compute-intensive training phases: dense FFN, attention computation).

TPU 8t Training Paradigm Optimization

OptimizationDetails
MoE TrainingNative Expert Parallel support (DeepSeek / Mixtral style)
Long-context TrainingOptimized for 1M+ token context training
RLHF / Post-trainingNative optimization for Online RL (DPO / PPO / GRPO)
Multimodal TrainingVision-language joint training (ViT + LLM synchronized)
AXIOMArm Axion CPU co-processing (data preprocessing / weight initialization)

Arm Axion CPU Integration

ItemSpecification
ArchitectureArm Neoverse V2 (64-core)
TDP~100 W
RoleHost CPU + Data loading + Preprocessing + Inference scheduling
SignificanceGoogle's custom Arm CPU enters TPU nodes for the first time

Axion integration = TPU nodes evolving towards "super nodes": TPU 8t is no longer a pure accelerator, but a TPU + Axion CPU co-processing system, competing with NVIDIA Vera CPU.

ScenarioRecommended Configuration
Gemini 3 TrainingTPU 8t pod 9,216 chips (single pod can train frontier models)
Llama 4 TrainingTPU 8t pod (hundred-billion-scale models)
Multimodal TrainingTPU 8t + Vision Transformer
Scientific ComputingTPU 8t + JAX 0.5+
RLHF Post-trainingTPU 8t (natively optimized)

Software Ecosystem

  • JAX 0.5+: Google's primary training framework
  • PyTorch/XLA 2.5+: PyTorch compatibility
  • TensorFlow 2.17+: Legacy framework
  • Paxml / Orbax: Google internal LLM training stack
  • MaxText: Google reference implementation
  • vLLM 0.8+ (experimental): Inference support

Use Cases

  • Frontier model training (Gemini 3/4, Anthropic, external customers)
  • MoE large-model training (native support)
  • Long-context training (1M+ token)
  • Multimodal training (ViT + LLM)
  • ❌ Inference scenarios (use TPU 8i instead of 8t)
  • ❌ Non-Google Cloud deployments

Vendor Information

ItemDetails
VendorGoogle Cloud
First Announced2026-04-22 (Google Cloud Next 2026)
Product Pagehttps://cloud.google.com/tpu
Cloud DeploymentGoogle Cloud only
CodenameTrillium 2