SemiAnalysis报道称，Anthropic正请求Google为其专用的TPUv9i SKU放弃Boardfly和Ocean Colour Scene。此举将通过消除路由核心，并与其他加速器（如Trainium或GPU）集成，从而为计算或内存腾出更多芯片空间。提议的网络架构涉及一个单层扁平化网络，包含多个交换机层，支持多达1,280个互连的TPU，并具有4:1的过载比。 AI
影响 可能提高AI加速器的计算密度和效率，影响未来的硬件设计。
排序理由 该集群包含来自SemiAnalysis的推文，讨论关于Google TPU架构的传言和潜在变化以及Anthropic的请求，而不是来自前沿实验室的直接公告。
AI 生成摘要 · Google Gemini · 来自 7 个来源。 我们如何撰写摘要 →
完整方法见我们的编辑标准。
To learn more, please check our networking model (7/7) https://t.co/kqt0IiSiFi
The main benefit for Anthropic would be more die space for compute or memory since the routing cores could be dropped as it should be handled by the switches, and a unified network architecture with other accelerators, such as Trainium or GPUs, as only Google’s TPUs are currently
This would leave 64 ports for cross-plane or intra-plane dragonfly interconnect, meaning a 4:1 oversubscription ratio and a maximum cluster size of up to 1,280 interconnected TPUs. (5/7)
Looking a the topology, we expect it to be a 1-tier flattened network, with 5 planes of 24 switches or 24 planes of 5 switches. Each switch should connect to 4 compute racks and each rack should connect to 24 switches at 300G per lane. (4/7) https://t.co/0n619Dwisy
Rumors have hinted towards an updated version of ALAB’s PCIe Scorpio-X switch featuring 320-radix. This updated version should bring 336G SerDes, opposed to the current 128G and could ultimately be manufactured by multiple vendors. (3/7) https://t.co/UCmuJmYCSe
This request seems to be coming from Anthropic, requesting Google to drop Boardfly and OCS for the AI lab’s dedicated TPUv9i SKU. (2/7)
Following 6D-Torus, there have been some industry chatters around a full electrical switching ICI for Triggerfish, Google’s next generation TPUv9i manufactured by MediaTek. (1/7)🧵 https://t.co/OXNpDSqZUw