Organízate con las colecciones
Guarda y clasifica el contenido según tus preferencias.
En esta página se proporciona información general sobre cómo funcionan las GPUs con Dataflow, incluidos los requisitos previos y los tipos de GPU admitidos.
Usar GPUs en tareas de Dataflow te permite agilizar algunas tareas de procesamiento de datos. Las GPUs pueden realizar determinados cálculos más rápido que las CPUs. Estos cálculos suelen ser numéricos o de álgebra lineal, y se utilizan a menudo en casos prácticos de procesamiento de imágenes y aprendizaje automático. El grado de mejora del rendimiento varía en función del caso práctico, el tipo de cálculo y la cantidad de datos procesados.
Requisitos previos para usar GPUs en Dataflow
Para usar GPUs con tu tarea de Dataflow, debes usar Runner v2.
Dataflow ejecuta el código de usuario en máquinas virtuales de trabajador dentro de un contenedor Docker.
Estas VMs de trabajador ejecutan Container-Optimized OS.
Para que las tareas de Dataflow usen GPUs, debes cumplir los siguientes requisitos:
Los controladores de GPU están instalados en las VMs de trabajador y el contenedor de Docker puede acceder a ellos. Para obtener más información, consulta Instalar controladores de GPU.
Para obtener información sobre las regiones y zonas en las que están disponibles las GPUs, consulta el artículo Disponibilidad de regiones y zonas de GPUs en la documentación de Compute Engine.
Cargas de trabajo recomendadas
En la siguiente tabla se ofrecen recomendaciones sobre el tipo de GPU que se debe usar para diferentes cargas de trabajo. Los ejemplos de la tabla son solo sugerencias y debes hacer pruebas en tu propio entorno para determinar el tipo de GPU adecuado para tu carga de trabajo.
Para obtener información más detallada sobre el tamaño de la memoria de la GPU, la disponibilidad de las funciones y los tipos de cargas de trabajo ideales para los distintos modelos de GPU, consulta la tabla comparativa general de la página de plataformas de GPU.
[[["Es fácil de entender","easyToUnderstand","thumb-up"],["Me ofreció una solución al problema","solvedMyProblem","thumb-up"],["Otro","otherUp","thumb-up"]],[["Es difícil de entender","hardToUnderstand","thumb-down"],["La información o el código de muestra no son correctos","incorrectInformationOrSampleCode","thumb-down"],["Me faltan las muestras o la información que necesito","missingTheInformationSamplesINeed","thumb-down"],["Problema de traducción","translationIssue","thumb-down"],["Otro","otherDown","thumb-down"]],["Última actualización: 2025-09-10 (UTC)."],[[["\u003cp\u003eDataflow jobs using GPUs can accelerate data processing, especially for numeric or linear algebra computations like those in image processing and machine learning.\u003c/p\u003e\n"],["\u003cp\u003eUsing GPUs in Dataflow requires Dataflow Runner v2 and incurs charges detailed on the Dataflow pricing page.\u003c/p\u003e\n"],["\u003cp\u003ePrerequisites for GPU usage include having GPU drivers installed on worker VMs and GPU libraries installed in the custom container image.\u003c/p\u003e\n"],["\u003cp\u003eDataflow supports several NVIDIA GPU types, including L4, A100 (40 GB and 80 GB), Tesla T4, P4, V100, and P100, each suited for different workload sizes and types.\u003c/p\u003e\n"],["\u003cp\u003eThe boot disk size for GPU containers should be increased to at least 50 gigabytes to prevent running out of disk space, due to the large nature of these containers.\u003c/p\u003e\n"]]],[],null,["\u003cbr /\u003e\n\n| **Note:** The following considerations apply to this GA offering:\n|\n| - Jobs that use GPUs incur charges as specified in the Dataflow [pricing page](/dataflow/pricing).\n| - To use GPUs, your Dataflow job must use [Dataflow Runner v2](/dataflow/docs/runner-v2).\n\n\u003cbr /\u003e\n\nThis page provides background information on how GPUs work with\nDataflow, including information about prerequisites and supported\nGPU types.\n\nUsing GPUs in Dataflow jobs lets you accelerate\nsome data processing tasks. GPUs can perform certain computations faster\nthan CPUs. These computations are usually numeric or linear algebra,\noften used in image processing and machine learning use cases. The\nextent of performance improvement varies by the use case, type of computation,\nand amount of data processed.\n\nPrerequisites for using GPUs in Dataflow\n\n\n- To use GPUs with your Dataflow job, you must use Runner v2.\n- Dataflow runs user code in worker VMs inside a Docker container. These worker VMs run [Container-Optimized OS](/container-optimized-os/docs). For Dataflow jobs to use GPUs, you need the following prerequisites:\n - GPU drivers are installed on worker VMs and accessible to the Docker container. For more information, see [Install GPU drivers](/dataflow/docs/gpu/use-gpus#drivers).\n - GPU libraries required by your pipeline, such as [NVIDIA CUDA-X libraries](https://developer.nvidia.com/gpu-accelerated-libraries) or the [NVIDIA CUDA Toolkit](https://developer.nvidia.com/cuda-toolkit), are installed in the custom container image. For more information, see [Configure your container image](/dataflow/docs/gpu/use-gpus#container-image).\n- Because GPU containers are typically large, to avoid [running out of disk space](/dataflow/docs/guides/common-errors#no-space-left), increase the default [boot disk size](/dataflow/docs/reference/pipeline-options#worker-level_options) to 50 gigabytes or more.\n\n\u003cbr /\u003e\n\nPricing\n\nJobs using GPUs incur charges as specified in the Dataflow\n[pricing page](/dataflow/pricing).\n\nAvailability **Note:** TPUs are also supported with Dataflow. For more information, see [Dataflow support for TPUs](/dataflow/docs/tpu/tpu-support).\n\nThe following GPU types are supported with Dataflow:\n\n| GPU type | `worker_accelerator` string |\n|---------------------|-----------------------------|\n| NVIDIA® L4 | `nvidia-l4` |\n| NVIDIA® A100 40 GB | `nvidia-tesla-a100` |\n| NVIDIA® A100 80 GB | `nvidia-a100-80gb` |\n| NVIDIA® Tesla® T4 | `nvidia-tesla-t4` |\n| NVIDIA® Tesla® P4 | `nvidia-tesla-p4` |\n| NVIDIA® Tesla® V100 | `nvidia-tesla-v100` |\n| NVIDIA® Tesla® P100 | `nvidia-tesla-p100` |\n| NVIDIA® H100 | `nvidia-h100-80gb` |\n| NVIDIA® H100 Mega | `nvidia-h100-mega-80gb` |\n\nFor more information about each GPU type, including performance data, see\n[Compute Engine GPU platforms](/compute/docs/gpus).\n\nFor information about available regions and zones for GPUs, see\n[GPU regions and zones availability](/compute/docs/gpus/gpu-regions-zones)\nin the Compute Engine documentation.\n\nRecommended workloads\n\nThe following table provides recommendations for which type of GPU to use for\ndifferent workloads. The examples in the table are suggestions only, and you\nneed to test in your own environment to determine the appropriate GPU type for\nyour workload.\n\nFor more detailed information about GPU memory size, feature availability, and\nideal workload types for different GPU models, see the\n[General comparison chart](/compute/docs/gpus#general_comparison_chart)\non the GPU platforms page.\n\n| Workload | A100, H100 | L4 | T4 |\n|------------------------|-------------|-------------|-------------|\n| Model fine tuning | Recommended | | |\n| Large model inference | Recommended | Recommended | |\n| Medium model inference | | Recommended | Recommended |\n| Small model inference | | Recommended | Recommended |\n\nWhat's next\n\n- See an example of a [developer workflow for building pipelines that use GPUs](/dataflow/docs/gpu/develop-with-gpus).\n- Learn how to [run an Apache Beam pipeline on Dataflow with GPUs](/dataflow/docs/gpu/use-gpus).\n- Work through [Processing Landsat satellite images with GPUs](/dataflow/docs/samples/satellite-images-gpus)."]]
