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
| Item | Specification |
|---|
| Architecture | TPU v6e (Trillium) |
| BF16 Compute | 918 TFLOPS (4.7× v5e) |
| INT8 Compute | 1,836 TOPS |
| HBM Capacity | 32 GB (2× v5e) |
| HBM Bandwidth | 1,638 GB/s |
| ICI Interconnect Bandwidth | 800 GB/s (bidirectional) |
| ICI Ports | 4 |
| DCN Bandwidth | 100 Gbps (2× v5e) |
| Pod Size | 256 chips (2D Torus) |
| vCPU (4-chip VM) | 180 |
| DRAM (4-chip VM) | 720 GB |
| Availability | Google Cloud only |
Trillium vs v5p Comparison
| Metric | v5p | v6e (Trillium) | Change |
|---|
| BF16 Compute | 459 TFLOPS | 918 TFLOPS | 2× |
| HBM Capacity | 95 GB | 32 GB | 1/3 |
| HBM Bandwidth | 2,575 GB/s | 1,638 GB/s | 0.64× |
| Pod Size | 8,960 | 256 | Smaller |
| Interconnect | 3D Torus | 2D Torus | Simplified |
| Efficiency | 1× | +67% | Improved |
Note: Trillium is better suited for mid-scale training and inference; v5p excels at ultra-large scale.
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