Google TPU 8i(推理)Google TPU 8i (inference)
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 里负责推理的是哪颗、为什么它更适合 agent 的低延迟场景」。8i 把单颗 HBM 提到 288GB、片上 SRAM 提到 384MB(前代 3 倍),配 19.2 Tb/s ICI 与 Boardfly 拓扑,把最大网络直径砍半;CAE 卸载全局通信,官方称片上延迟最多降 5 倍。单 pod 1,152 颗、FP8 11.6 EFLOPS、HBM 合计 331.8TB;官方称每美元性能提升 80%。
不主张:官方发布页没有给单颗 FP4 算力(行业转述为 10.1 PF,本卡不写);11.6 EFLOPS 是 FP8 pod 合计,不是单颗;80% 与 5× 都是 Google 口径,不是独立实测;没有公开价格。Answers which chip handles inference in Google's eighth-generation TPU line, and why it suits low-latency agent workloads. TPU 8i raises per-chip HBM to 288 GB and on-chip SRAM to 384 MB (3x the previous generation), pairs 19.2 Tb/s ICI with the Boardfly topology and more than halves the maximum network diameter, while the CAE offloads collective operations; Google says on-chip latency falls by up to 5x. A pod holds 1,152 chips, 11.6 FP8 exaflops and 331.8 TB of HBM; Google claims 80% better performance per dollar.
It does not claim: the launch page gives no per-chip FP4 figure (third-party reporting says 10.1 PF and is not written here); 11.6 EFLOPS is an FP8 pod total, not a single chip; the 80% and 5x figures are Google's, not independent measurements; no public pricing.
关键参数Key specs
- 定位 / role
- 低延迟推理、采样与 agent 服务
- 单颗 HBM / HBM per chip
- 288 GB
- 单颗 SRAM / SRAM per chip
- 384 MB(前代 3×)
- Pod 规模 / pod size
- 1,152 chips
- Pod 算力 / pod compute
- 11.6 EFLOPS FP8
- Pod HBM 合计 / HBM per pod
- 331.8 TB
- Scale-up 带宽 / scale-up
- 19.2 Tb/s per chip
- 主机 / host
- Google Axion(Arm)
- 网络 / network
- Boardfly
- 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.
TPU 8i pairs 288 GB of high-bandwidth memory with 384 MB of on-chip SRAM — 3x more than the previous generation — keeping a model's active working set entirely on-chip.
we doubled the Interconnect (ICI) bandwidth to 19.2 Tb/s. Our new Boardfly architecture reduces the maximum network diameter by more than 50%
Our new on-chip Collectives Acceleration Engine (CAE) offloads global operations, reducing on-chip latency by up to 5x
These innovations deliver 80% better performance-per-dollar compared to the previous generation
Total HBM capacity per pod (TB) 331.8
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
- 在线服务Serving
- 价格口径Price basis
- 未公布单价No public price · not-published