[[["易于理解","easyToUnderstand","thumb-up"],["解决了我的问题","solvedMyProblem","thumb-up"],["其他","otherUp","thumb-up"]],[["很难理解","hardToUnderstand","thumb-down"],["信息或示例代码不正确","incorrectInformationOrSampleCode","thumb-down"],["没有我需要的信息/示例","missingTheInformationSamplesINeed","thumb-down"],["翻译问题","translationIssue","thumb-down"],["其他","otherDown","thumb-down"]],["最后更新时间 (UTC):2025-08-19。"],[],[],null,["# TPU v5p\n=======\n\nThis document describes the architecture and supported configurations of\nCloud TPU v5p.\n\nSystem architecture\n-------------------\n\nThis section describes the system architecture specific to the v5p version. Each\nTensorCore has four Matrix Multiply Units (MXU), a vector unit, and a scalar\nunit.\n\nThere are 8960 chips in a single v5p slice. The largest job that can be scheduled\nis a 96 cube (6144 chip) job.\n\nThe following table shows the key specifications for a v5p.\n\n| **Important:** All 4x4x4 and larger slices (one cube) have full 3D torus connectivity. Slices smaller than a full cube are 3D connected, however, they don't have wrap-around links that make them a 3D torus.\n\nConfigurations\n--------------\n\nA TPU v5p Pod is composed of 8960 chips interconnected with reconfigurable\nhigh-speed links. TPU v5p's flexible networking lets you connect the chips in a\nsame-sized slice in multiple ways. When you create a TPU slice using the\n`gcloud compute tpus tpu-vm create` command, you specify its type and shape\nusing the [`AcceleratorType`](#using-accelerator-type) parameter.\n\nThe following table shows the most common single-slice shapes supported with v5p,\nplus most (but not all) full cube shapes greater than 1 cube. The maximum v5p\nshape is 16x16x24 (6144 chips, 96 cubes).\n| **Note:** All v5p TPU VMs use the `ct5p-hightpu-4t` machine type.\n\nSingle slice training is supported for up to 6144 chips. You can scale up to\n18432 chips using Multislice. For more information about\nMultislice, see [Cloud TPU Multislice Overview](/tpu/docs/multislice-introduction).\n\n### Using the AcceleratorType parameter\n\nWhen you allocate TPU resources, you use the `--accelerator-type` argument to\nspecify the number of TensorCores in a slice. `--accelerator-type` is a\nformatted string \"**v** `$VERSION_NUMBER`**p** `-$CORES_COUNT`\".\nFor example, `v5p-32` specifies a v5p TPU slice with 32 TensorCores (16 chips).\n\nTo provision TPUs for a v5p training job, use one of the following accelerator\ntypes in your CLI or TPU API creation request:\n\n- v5p-8\n- v5p-16\n- v5p-32\n- v5p-64\n- v5p-128 (one full cube/rack)\n- v5p-256 (2 cubes)\n- v5p-512\n- v5p-1024 ... v5p-12288\n\nThe following command creates a v5p TPU slice with 256 v5p TensorCores (128 chips)\nfor training: \n\n```bash\n $ gcloud compute tpus tpu-vm create your-tpu-name \\\n --zone=us-east5-a \\\n --accelerator-type=v5p-256 \\\n --version=v2-alpha-tpuv5\n```\n\nFor more information about managing TPUs, see [Manage TPUs](/tpu/docs/managing-tpus-tpu-vm).\nFor more information about the system architecture of Cloud TPU, see\n[System architecture](/tpu/docs/system-architecture).\n\n### Cloud TPU ICI resiliency\n\nICI resiliency helps improve fault tolerance of optical links and optical\ncircuit switches (OCS) that connect TPUs between cubes. (ICI connections within\na cube use copper links that are not impacted). ICI resiliency allows ICI\nconnections to be routed around OCS and optical ICI faults. As a result, it\nimproves the scheduling availability of TPU slices, with the trade-off of\ntemporary degradation in ICI performance.\n\nSimilar to Cloud TPU v4, ICI resiliency is enabled by default for v5p slices\nthat are one cube or larger:\n\n- v5p-128 when specifying accelerator type\n- 4x4x4 when specifying accelerator config\n\n### VM, host and slice properties\n\nRelationship between the number of TensorCores, chips, hosts/VMs, and cubes\nin a Pod:"]]