以下部分引用 AI 总结,现阶段 AI 仍然有幻觉。请以内容的原文为准。
数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15
Google TPU v8
19 atoms · 跨 5 天 · 首见 2026-04-20 · 最近 2026-05-24
三色: 🟦 fact 19 · 🟥 take 0 · stance ▲3/▼0/◆4
来源: X 19
时态: fresh:19
标签: 好信源:9 · 好数字:9 · 好观点:1
事实 (fact) (19)
- 2026-05-24 · requires above 600,000 passive components per rack (compute 400,000 + power 200,000)
compute node demand is estimated at about 400,000 units and power distribution usage at about 200,000, pushing total per-rack passive component demand above 600,000 units
tags: 好数字 · → daily
value: qty=600,000 units · date=
- 2026-04-23 · 大规模集群配套CPU架构
Google stated that TPU v8 clusters will largely be supported by their self-designed Axion ARM-based CPUs.
tags: 好信源 · → daily
value: direction=ARM
- 2026-04-22 · 训练性价比提升
谷歌第八代TPU对比上一代Ironwood:训练性jia比提升2.7倍
tags: 好数字 · → daily
value: qty=2.7x
- 2026-04-22 · 推理性价比提升
推理性jia比提升1.8倍
tags: 好数字 · → daily
value: qty=1.8x
- 2026-04-22 · 能效提升
能效翻倍
tags: 好数字 · → daily
value: qty=2x
- 2026-04-22 · 集群规模最大
最多能扩到100万颗
tags: 好数字 · → daily
value: qty=100万颗
- 2026-04-22 · 集群起始规模
集群规模从数万颗扩到13.4万颗
tags: 好数字 · → daily
value: qty=13.4万颗
- 2026-04-22 · 全互联通信跳数减少
全互联通信从16跳降到7跳,减少56%
tags: 好数字 · → daily
value: qty=56%
- 2026-04-20 · 芯片型号
Google’s next-generation TPU v8 chips are expected to feature prominently at its Google Cloud Next event this week (April 22–24, 2026), according to media and supply-chain reports.
tags: 好信源 · → daily
- 2026-04-20 · 与Axion CPU集成
The chips are designed to integrate tightly with Google’s Axion ARM-based CPUs as part of its AI supercomputing architecture.
tags: 好信源 · → daily
- 2026-04-20 · 发布活动
Google’s next-generation TPU v8 chips are expected to feature prominently at its Google Cloud Next event this week (April 22–24, 2026)
tags: 好信源 · → daily
value: date=2026-04-22 to 2026-04-24
- 2026-04-20 · 型号名称
It said no, v8t and v8i are correct.
tags: 好信源 · → daily
- 2026-04-24 · networking topology
Two scale-up networking topologies
tags: 好信源 · → daily
- 2026-04-24 · splits into two types
TPU splits in two: 8t training, 8i inference.
tags: 好信源 · → daily
- 2026-04-24 · Virgo scale-out fabric bandwidth
Virgo: 47 Pb/s scale-out fabric, 100% OCS.
tags: 好数字 · → daily
value: qty=47 Pb/s
- 2026-04-24 · Boardfly scale-up topology
Boardfly scale-up: copper PCB + AECs inside racks, OCS between groups.
tags: 好信源 · → daily
- 2026-04-24 · training uses 3D torus topology
Training uses 3D torus (Rubik's Cube).
tags: 好信源 · → daily
- 2026-04-22 · ratio of TPU per CPU changed
TPU v8 swaps to Axiom and 2 TPU per CPU. Previously it was 4 TPU per CPU
tags: 好数字 · → daily
value: qty=2 TPU per CPU · date=v8 generation
- 2026-04-24 · inference topology different from training
Inference doesn't. Workload-specific topologies now.
tags: 好观点 · → daily
时间轴 (近 20)
- 2026-05-24 ·
fact · requires above 600,000 passive components per rack (compute · → daily
- 2026-04-24 ·
fact · networking topology · → daily
- 2026-04-24 ·
fact · splits into two types · → daily
- 2026-04-24 ·
fact · Virgo scale-out fabric bandwidth · → daily
- 2026-04-24 ·
fact · Boardfly scale-up topology · → daily
- 2026-04-24 ·
fact · training uses 3D torus topology · → daily
- 2026-04-24 ·
fact · inference topology different from training · → daily
- 2026-04-23 ·
fact · 大规模集群配套CPU架构 · → daily
- 2026-04-22 ·
fact · 训练性价比提升 · → daily
- 2026-04-22 ·
fact · 推理性价比提升 · → daily
- 2026-04-22 ·
fact · 能效提升 · → daily
- 2026-04-22 ·
fact · 集群规模最大 · → daily
- 2026-04-22 ·
fact · 集群起始规模 · → daily
- 2026-04-22 ·
fact · 全互联通信跳数减少 · → daily
- 2026-04-22 ·
fact · ratio of TPU per CPU changed · → daily
- 2026-04-20 ·
fact · 芯片型号 · → daily
- 2026-04-20 ·
fact · 与Axion CPU集成 · → daily
- 2026-04-20 ·
fact · 发布活动 · → daily
- 2026-04-20 ·
fact · 型号名称 · → daily
← 实体目录 · 系统日志