Silicon & Pod Architecture
Google TPU v6, codenamed Trillium, is Google's sixth-generation Tensor Processing Unit tailored exclusively for training and serving massive frontier models like Gemini. Trillium achieves a 4.7x increase in peak compute per chip over TPU v5e while doubling HBM bandwidth and inter-chip interconnect (ICI).
| Specification | TPU v6 (Trillium) | TPU v5p | TPU v5e |
|---|---|---|---|
| Peak Compute Metric | 4.7x vs v5e (High FP8/BF16) | 459 TFLOPS (BF16) | 197 TFLOPS (BF16) |
| Memory Architecture | High-Density HBM | 95 GB HBM2e (2.76 TB/s) | 16 GB HBM2e (819 GB/s) |
| Pod Scalability | Up to tens of thousands | 8,960 chips in 3D Torus | 256 chips |
| Network Topology | 3D Torus + Optical Circuit Switching (OCS) | 3D Torus with OCS | 2D Torus |
| Power Efficiency | 67% improvement per Watt | High-efficiency custom Pod | Cost-optimized |
Optical Circuit Switching (OCS) Advantage
Unlike traditional spine-leaf Ethernet or InfiniBand networks requiring thousands of expensive electrical switches, Google TPU Pods dynamically reconfigure 3D Torus interconnects via MEMS-based Optical Circuit Switches. This eliminates electrical packet routing overhead and enables fault recovery in seconds by bypassing failing nodes optically.