Skip to main content

Google Cloud TPU v6e (Trillium)

Overview

Google TPU v6e (codenamed Trillium) reached GA in December 2024, marking Google's 6th-generation TPU. It delivers 4.7× the peak compute of v5e, 918 TFLOPS BF16 compute, doubled HBM capacity and ICI bandwidth. It was used to train Gemini 2.0. Trillium is a core component of the Google Cloud AI Hypercomputer architecture, scalable to 100,000+ chips via the Jupiter network.

Core Specifications

ItemSpecification
ArchitectureTPU v6e (Trillium)
BF16 Compute918 TFLOPS (4.7× v5e)
INT8 Compute1,836 TOPS
HBM Capacity32 GB (2× v5e)
HBM Bandwidth1,638 GB/s
ICI Interconnect Bandwidth800 GB/s (bidirectional)
ICI Ports4
DCN Bandwidth100 Gbps (2× v5e)
Pod Size256 chips (2D Torus)
vCPU (4-chip VM)180
DRAM (4-chip VM)720 GB
AvailabilityGoogle Cloud only

Trillium vs v5p Comparison

Metricv5pv6e (Trillium)Change
BF16 Compute459 TFLOPS918 TFLOPS
HBM Capacity95 GB32 GB1/3
HBM Bandwidth2,575 GB/s1,638 GB/s0.64×
Pod Size8,960256Smaller
Interconnect3D Torus2D TorusSimplified
Efficiency+67%Improved

Note: Trillium is better suited for mid-scale training and inference; v5p excels at ultra-large scale.

Vendor Information

ItemDetails
ManufacturerGoogle LLC
Official Websitehttps://cloud.google.com/tpu
Product Pagehttps://cloud.google.com/tpu/docs/v6e
ReleaseDecember 2024 GA
TrainedGemini 2.0

Key Features

  • SparseCore 3rd Gen: 2× performance for embedding-intensive models
  • Training/Inference fungible: Same quota for training and inference
  • Multislice: Scalable to 100,000+ chips
  • Multi-host inference: Supports 70B+ models
  • XLA compiler optimization: First-class JAX/PyTorch/TF support

Use Cases

  • LLM training (Gemini 2.0, PaLM)
  • LLM inference
  • Multimodal models
  • Embedding-intensive models (DLRM)
  • Diffusion model inference