TPU v6e
This document describes the architecture and supported configurations of Cloud TPU v6e (Trillium). On all technical surfaces, such as the API and logs, and throughout this document, Trillium will be referred to as v6e.
With a 256-chip footprint per Pod, v6e shares many similarities with v5e. This system is optimized to be the highest value product for transformer, text-to-image, and convolutional neural network (CNN) training, fine-tuning, and serving.
System architecture
Each v6e chip contains one TensorCore. Each TensorCore has 2 matrix-multiply units (MXU), a vector unit, and a scalar unit. The following table shows the key specifications and their values for TPU v6e compared to TPU v5e.
| Specification | v5e | v6e |
|---|---|---|
| Performance/total cost of ownership (TCO) (expected) | 0.65x | 1 |
| Peak compute per chip (bf16) | 197 TFLOPs | 918 TFLOPs |
| Peak compute per chip (Int8) | 393 TOPs | 1836 TOPs |
| HBM capacity per chip | 16 GB | 32 GB |
| HBM bandwidth per chip | 800 GBps | 1600 GBps |
| Inter-chip interconnect (ICI) bandwidth | 1600 Gbps | 3200 Gbps |
| ICI ports per chip | 4 | 4 |
| DRAM per host | 512 GiB | 1536 GiB |
| Chips per host | 8 | 8 |
| TPU Pod size | 256 chips | 256 chips |
| Interconnect topology | 2D torus | 2D torus |
| BF16 peak compute per Pod | 50.63 PFLOPs | 234.9 PFLOPs |
| All-reduce bandwidth per Pod | 51.2 TB/s | 102.4 TB/s |
| Bisection bandwidth per Pod | 1.6 TB/s | 3.2 TB/s |
| Per-host NIC configuration | 2 x 100 Gbps NIC | 4 x 200 Gbps NIC |
| Data center network bandwidth per Pod | 6.4 Tbps | 25.6 Tbps |
| Special features | - | SparseCore |
Supported configurations
The following table shows the 2D slice shapes that are supported for v6e:
| Topology | TPU chips | Hosts | VMs | Machine type (GKE API) | Scope |
|---|---|---|---|---|---|
| 1x1 | 1 | 1/8 | 1 | ct6e-standard-1t |
Sub-host |
| 2x2 | 4 | 1/2 | 1 | ct6e-standard-4t |
Sub-host |
| 2x4 | 8 | 1 | 1 | ct6e-standard-8t |
Single-host |
| 2x4 | 8 | 1 | 2 | ct6e-standard-4t |
Single-host |
| 4x4 | 16 | 2 | 4 | ct6e-standard-4t |
Multi-host |
| 4x8 | 32 | 4 | 8 | ct6e-standard-4t |
Multi-host |
| 8x8 | 64 | 8 | 16 | ct6e-standard-4t |
Multi-host |
| 8x16 | 128 | 16 | 32 | ct6e-standard-4t |
Multi-host |
| 16x16 | 256 | 32 | 64 | ct6e-standard-4t |
Multi-host |
Slices with 8 chips (v6e-8) attached to a single VM are optimized for
inference, allowing all 8 chips to be used in a single serving workload. You can
perform multi-host inference using Pathways on Cloud. For more information, see
Perform multihost inference using Pathways.
For information about the number of VMs for each topology, see VM Types.
VM types
Each TPU v6e VM can contain 1, 4, or 8 chips. 4-chip and smaller slices have the same non-uniform memory access (NUMA) node. For more information about NUMA nodes, see Non-uniform memory access on Wikipedia.
v6e slices are created using half-host VMs, each with 4 TPU chips. There are two exceptions to this rule:
v6e-1: A VM with only a single chip, primarily intended for testingv6e-8: A full-host VM that has been optimized for an inference use case with all 8 chips attached to a single VM.
The following table shows a comparison of TPU v6e VM types:
| VM type | Number of vCPUs per VM | RAM (GB) per VM | Number of NUMA nodes per VM |
|---|---|---|---|
| 1-chip VM | 44 | 176 | 1 |
| 4-chip VM | 180 | 720 | 1 |
| 8-chip VM | 180 | 1440 | 2 |