Google 推出了其新的 TPUv8t（一种专注于训练的加速器）以及一个名为 Virgo 的新颖网络架构。该新系统实现了显著的代际飞跃，与前代产品相比，每个加速器的带宽提高了四倍，延迟降低了 40%。Virgo 旨在互连多达 134,400 颗芯片，支持总计 47 PB/s 的无阻塞带宽。 AI
影响 TPU 和网络架构的这一进步可以显著提高 Google 内部工作负载和云服务的 AI 训练效率和规模。
排序理由 该集群详细介绍了主要科技公司的新硬件产品发布和重要的基础设施公告。
AI 生成摘要 · Google Gemini · 来自 4 个来源。 我们如何撰写摘要 →
To learn more, please check our networking model: https://t.co/kqt0IiSiFi (4/4)
Within the plane, each switch is connected in full mesh using 800G 2xFR4 OSFPs in 8x100G breakout mode on the L1 layer and 400G FR4 QSFPs in 4x100G breakout mode on the L2 layer. To maintain a non-blocking topology, Google is doubling some links between the L1 and the L2 layers.
This marks a generational leap with up to 4x the bandwidth per accelerator and a 40% lower latency compared to previous generation. On a network perspective, this architecture connects 14 TPU pods (9,600 interconnected TPUs each, via 3D Torus ICI) on a flat 2-layers topology, htt…
With the introduction of the TPUv8t, their new training focused TPU, Google unveiled a new scale-out network architecture called Virgo. Virgo is able to interconnect up to 134,400 chips with up to 47 Pbps of non-blocking bi-sectional bandwidth. (1/4)🧵 https://t.co/J4TCJ70BFP