算力服务

高性能AI算力池是依托商汤多年研发与运营超大规模AI超级计算集群经验，打造的高性能、弹性扩展、安全稳定、智能调度的AI算力平台，大幅提高算力资源的利用率及AI应用的效率，助力AI业务增长。产品功能支持多种框架的单机和分布式任务的生命周期管理支持TensorFlow、PyTorch训练框架、MPI方式的单机和分布式训练任务以及各种角色的配置，支持按需创建训练任务以及管理任务各种状态操作；提供多种调度策略满足各种任务的高效低成本诉求；支持RDMA网络通过 RDMA（IB/RoCE）或 TCP/IP 网络连接各个节点，实现资源池弹性扩展；支持任务的可观测性提供多种任务资源、业务指标的监控和日志服务，满足算法工程师在调试过程中运维诉求。

The High-Performance AI Computing Pool is a high-performance, elastically scalable, secure and stable, intelligently scheduled AI computing platform built based on SenseTime’s years of experience in R&D and operation of ultra-large-scale AI supercomputing clusters. It greatly improves the utilization rate of computing resources and the efficiency of AI applications, and boosts the growth of AI businesses. Its functions support lifecycle management of single-node and distributed tasks across multiple frameworks: it supports single-node and distributed training tasks using TensorFlow, PyTorch training frameworks and MPI, as well as configuration of various roles, and supports on-demand creation of training tasks and operations for managing various task states; it provides multiple scheduling strategies to meet the efficient and low-cost requirements of various tasks; it supports RDMA network, connecting various nodes via RDMA (IB/RoCE) or TCP/IP networks to enable elastic expansion of the resource pool; it supports task observability by providing monitoring and logging services for various task resources and business indicators, meeting the operation and maintenance needs of algorithm engineers during debugging.

搜集汇总

背景与挑战

背景概述

该数据集描述了商汤科技提供的高性能AI算力服务，包括支持多种框架的分布式任务管理、RDMA网络支持以及任务可观测性等功能，适用于互联网/IT/电子/通信行业领域。