跳到主要内容Skip to main content

← 返回硬件算力← Back to Compute

加速卡Accelerators

Google TPU v5eGoogle TPU v5e

回答「v5e 单芯片的 BF16 算力、HBM、ICI 与 Pod 规模」。197 TFLOPS 是 BF16 峰值;16GB HBM2 / 819 GB/s 是单芯片;256 芯片是 Pod 上限。
不主张:不是 v5p 或 v6e;不是单卡零售产品;实际训练/推理吞吐取决于拓扑、批量与框架。
Answers per-chip BF16 compute, HBM, ICI and Pod size. 197 TFLOPS is BF16 peak; 16GB HBM2 / 819 GB/s is per chip; 256 chips is the Pod ceiling.
It does not claim: this is not v5p or v6e; it is not a retail add-in card; real throughput depends on topology, batching and framework.

关键参数Key specs

BF16 算力
197 TFLOPS / 芯片
HBM2
16 GB / 芯片
HBM 带宽
819 GB/s / 芯片
ICI
400 GB/s / 芯片
Pod
256 芯片

官方原文怎么说What the source says

每个芯片的峰值计算能力 (bf16) | 197 TFLOPs
1 https://cloud.google.com/tpu/docs/v5e
HBM2 容量和带宽 | 16 GB,819 GBps
2 https://cloud.google.com/tpu/docs/v5e
双向芯片间互连 (ICI) 带宽(每芯片) | 400 GBps
3 https://cloud.google.com/tpu/docs/v5e
TPU Pod 大小 | 256 个芯片
4 https://cloud.google.com/tpu/docs/v5e

资料与许可Source and licence

资料时间Source checked
2026-11-19
许可 / 类型Licence / kind
official-spec-page
许可说明Licence evidence
https://cloud.google.com/tpu/docs/v5e
类目Category
加速卡Accelerators

去官网看原页Open the official page →

← 返回硬件算力← Back to Compute