Support for Heterogeneous Computing in All Scenarios

Elastic GPU Service provides a complete service system that combines software and hardware to help you flexibly allocate resources, elastically scale your system, improve computing power, and lower the cost of your AI-related business. It applies to scenarios (such as deep learning, video encoding and decoding, video processing, scientific computing, graphical visualization, and cloud gaming). Elastic GPU Service provides GPU-accelerated computing capabilities and ready-to-use, scalable GPU computing resources. GPUs have unique advantages in performing mathematical and geometric computing, especially floating-point and parallel computing. GPUs provide 100 times the computing power of their CPU counterparts.

Varied Computing Capabilities

It has a large number of arithmetic logic units (ALUs) that can be used for large-scale parallel computing. Elastic GPU Service uses the latest GPU acceleration chips and provides various accelerator cards (such as FPGA, GPU, and ASIC) to serve business purposes (such as AI, graphics, transcoding, and encryption).

Ease of Use

GPU resources are globally deployed across different geographical locations. Simple logic control units allow you to scale your system based on your business requirements. Elastic GPU Service also provides auxiliary tools (like AIACC, FastGPU, and cGPU).

High Network Performance

It uses the SHENLONG architecture to improve server performance and reduce I/O latency. GPU supports up to 24 million pps, a bandwidth of up to 64 Gbit/s over VPCs, and an 800G RDMA network. It is suitable for high-throughput scenarios where multiple threads run in parallel to process computing tasks.

GPU Software for Improving Computing Efficiency

AIACC-Training Alibaba Cloud AIACC-Training is an AI accelerator optimized for Alibaba Cloud environments. It can significantly improve the efficiency of AI distributed training and network bandwidth utilization. AIACC-Training has set two world records: Fastest training speed in the DAWNBench ImageNet competition (held by Stanford University) Lowest training cost in the DAWNBench ImageNet competition (held by Stanford University) Features

•  Supports Main Frameworks Distributed training frameworks: TensorFlow, PyTorch, MXNet, and Caffe
•  50%-300% Performance Improvements Bandwidth-intensive network models
•  High-Performance Communication for One or More Multi-GPU Servers Supports FP16 gradient compression and mixed precision compression
•  API Extensions for MXNet Supports data parallelism and model parallelism of the InsightFace type
•  Deep Optimization for RDMA Networks Hybrid link communication (RDMA and VPC) AIACC-Inference Alibaba Cloud AIACC-Inference is an AI accelerator optimized for Alibaba Cloud environments. It can significantly improve GPU utilization and inference performance. AIACC-Inference has set two world records: Lowest inference latency in the DAWNBench ImageNet competition (held by Stanford University) Lowest inference cost in the DAWNBench ImageNet competition (held by Stanford University) Features
•  Supports Multiple Frameworks Tensorflow, Pytorch, MXNet, and other deep learning frameworks can export models in the Open Neural Network Exchange (ONNX) format to improve inference performance
• 30%-400% Performance Improvements Compute-intensive network models
•  Supports Multiple Model Precisions Model optimization on FP32 and FP16 GPU Cluster Deployment Tool Alibaba Cloud FastGPU is a set of fast deployment tools for GPU clusters to help deploy GPU computing resources on the cloud with just a few clicks. FastGPU is simple to configure and can be used anywhere with ease. FastGPU provides a time-saving, cost-effective, and easy-to-use solution for the fast deployment of GPU clusters. Features
•  Quick Deployment API operations for fast deployment of offline training and inference scripts in GPU clusters
•  Easy Management Provides a command-line tool to manage the status and lifecycle of GPU clusters
•  Efficient and Time-Saving You do not need to perform deployment operations for computing, storage, and network at the IAAS layer of Alibaba Cloud. The appropriate environment is automatically attained when you obtain cluster resources. GPU Sharing Software for Containers It allows multiple containers to use a single GPU by splitting and assigning GPU resources to multiple isolated containers. cGPU can run multiple containers on a single GPU and isolate the GPU applications among the containers. This way, the GPU hardware resource utilization is improved. Features
•  GPU Splitting Splits GPU resources to improve GPU utilization
•  GPU Sharing Cost savings through GPU sharing across multiple AI applications
•  Flexibility Flexibly splits computing power and GPU memory to meet application requirements