加速卡Accelerators
Google TPU7x(Ironwood)Google TPU7x (Ironwood)
回答「第七代 TPU 单芯片给多少 HBM、FP8/BF16 算力,以及一个 Pod 有多大」。192GiB HBM3E、7.38TB/s、4614 TF FP8 是单芯片;9216 chips per pod 是 Pod 规模。它不是 GPU,不能按 CUDA 生态的选型逻辑直接套。
不主张:TPU 的 Pod 规模不是单卡可比的吞吐;4614 TF 是 FP8 峰值;不是所有框架都支持(官方写明 TensorFlow 不支持,JAX/PyTorch 支持);不是「比 GPU 更强」的结论。Answers per-chip HBM, FP8/BF16 compute and Pod size for the seventh-generation TPU. 192GiB HBM3E, 7.38TB/s and 4614 TF FP8 are per chip; 9,216 chips per Pod is the Pod footprint. It is not a GPU and cannot be slotted into CUDA-ecosystem selection logic.
It does not claim: Pod scale is not per-card throughput; 4614 TF is FP8 peak; not every framework is supported (the docs say TensorFlow is not supported, JAX/PyTorch are); it is not a claim that TPUs are stronger than GPUs.
关键参数Key specs
- HBM / chip
- 192 GiB HBM3E
- 带宽 / bandwidth
- 约 7.37 TB/s(7380 GBps)
- FP8
- 4614 TFLOPS / chip
- BF16
- 2307 TFLOPS / chip
- Pod 规模 / pod
- 9,216 chips
- ICI
- 1.2 TB/s 双向 / chip
- 框架 / frameworks
- JAX、PyTorch(TensorFlow 不支持)
官方原文怎么说What the source says
TPU7x is the first release within the Ironwood family, Google Cloud's seventh generation TPU.
HBM capacity per chip (GiB) 95 32 192
Peak compute per chip (FP8) (TFLOPs) 459 918 4614
With a 9,216-chip footprint per Pod
Bidirectional inter-chip interconnect (ICI) bandwidth per chip (GBps) 1200 800 1200
资料与许可Source and licence
- 资料时间Source checked
- 2026-11-18
- 许可 / 类型Licence / kind
- official-spec-page
- 许可说明Licence evidence
- https://docs.cloud.google.cn/tpu/docs/tpu7x
- 类目Category
- 加速卡Accelerators