跳到主要内容Skip to main content

← 返回硬件算力← Back to Compute

加速卡Accelerators未公布单价No public price

Google TPU 8t(训练)Google TPU 8t (training)

未公布单价No public price未公布单价No public price · 官方页未给公开单价;这类产品通常按整机、云实例或询价成交,不拿经销商价冒充官方价。No public unit price on the official page; these usually sell as systems, cloud instances or on request, so dealer quotes are not presented as official.

回答「Google 第八代 TPU 里负责训练的是哪颗、一个 pod 能到什么规模」。8t 与 8i 同代分叉:8t 走 3D Torus + Virgo,单 pod 9,600 颗、2PB 共享 HBM、121 EFLOPS FP4,官方称每 pod 算力接近前代 3 倍,目标把前沿模型开发周期从月压到周。
不主张:官方发布页只给 pod 级数字,没有给单颗 HBM 容量与单颗 FP4 算力;行业转述的 216GB / 128MB SRAM / 12.6 PF 是第三方口径,本卡不写。121 EFLOPS 是 FP4 pod 合计,不是单颗;9,600 颗是 8t pod 上限,不是单集群上限(官方另称 Virgo 可扩到百万颗逻辑集群)。没有公开价格与独立实测。
Answers which chip handles training in Google's eighth-generation TPU line, and how large a pod gets. TPU 8t and 8i split the generation: 8t uses 3D Torus plus Virgo, scales to 9,600 chips and 2 PB of shared HBM per pod, and delivers 121 FP4 exaflops; Google says that is nearly 3x the compute per pod over the previous generation and aims to cut frontier-model development from months to weeks.
It does not claim: the launch page publishes pod-level figures only, not per-chip HBM capacity or per-chip FP4 throughput; the widely repeated 216GB / 128MB SRAM / 12.6 PF figures are third-party reporting and are not written here. 121 EFLOPS is an FP4 pod total, not a single chip; 9,600 chips is the 8t pod size, not a single-cluster ceiling (Google separately says Virgo can scale to a million chips in one logical cluster). No public pricing or independent measurement.

关键参数Key specs

定位 / role
大规模预训练
Pod 规模 / pod size
9,600 chips
共享显存 / shared HBM
2 PB per pod
Pod 算力 / pod compute
121 EFLOPS FP4
Scale-up 带宽 / scale-up
19.2 Tb/s per chip
Scale-out 带宽 / scale-out
400 Gb/s per chip
主机 / host
Google Axion(Arm)
网络 / network
3D Torus + Virgo
状态 / status
2026-04-22 发布;官方称 2026 年内 GA

官方原文怎么说What the source says

Today at Google Cloud Next, we are introducing the eighth generation of Google's custom Tensor Processor Unit (TPU), coming soon with two distinct, purpose-built architectures for training and inference: TPU 8t and TPU 8i.
1 https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/eighth-generation-tpu-agentic-era/
A single TPU 8t superpod now scales to 9,600 chips and two petabytes of shared high bandwidth memory, with double the interchip bandwidth of the previous generation. This architecture delivers 121 ExaFlops of compute and allows the most complex models to leverage a single, massive pool of memory.
2 https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/eighth-generation-tpu-agentic-era/
TPU 8t is built to reduce the frontier model development cycle from months to weeks.
3 https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/eighth-generation-tpu-agentic-era/
delivers nearly 3x the compute performance per pod over the previous generation
4 https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/eighth-generation-tpu-agentic-era/
Scale-out networking bandwidth (Gb/s per chip) 400
5 https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/eighth-generation-tpu-agentic-era/

资料与许可Source and licence

资料时间Source checked
2026-11-20
许可 / 类型Licence / kind
official-spec-page
许可说明Licence evidence
https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/eighth-generation-tpu-agentic-era/
类目Category
加速卡Accelerators
状态Status
已发布 / 未出货Announced
适合Use cases
训练Training
价格口径Price basis
未公布单价No public price · not-published

同类目还有More in this category

下一步Next step

去官网看原页Open the official page →

← 返回硬件算力← Back to Compute