Parts/Accelerators/Google TPU v6e (Trillium)
Accelerators · v6e · 2024
Google TPU v6e (Trillium)
918 TFLOPS BF16 per chip, 32GB HBM, 256-chip pods on a 2D torus. Never sold as a part: it appears only as rented capacity.
Why this part
Trillium, at 918 TFLOPS of BF16 per chip in pods of 256 on a 2D torus. Google's torus topology is the quiet difference from an NVIDIA cluster: no switches, direct links between neighbours, cheaper to build and excellent for the regular communication patterns training produces. Like every TPU it is capacity rather than a product, and the only way to count it is to count what Google says it deployed.
Systems using this part
- Systems
- 3
- Rmax underneath
- 0.0 GFlop/s
- Units recorded
- 592,500
| System | Role | Country | Units | System Rmax | Supplied by | Rank |
|---|---|---|---|---|---|---|
| Google Council Bluffs East TPU Campus 2024-11 | Accelerator | USA | 125,200 | - | Alphabet | - |
| Google Columbus TPU Campus 2024-04 | Accelerator | USA | 204,600 | - | Alphabet | - |
| Google New Albany TPU Campus 2024-04 | Accelerator | USA | 262,700 | - | Alphabet | - |
“Units recorded” sums only the edges where a public source states a quantity. It is a floor, not a total, and should never be read as installed base.