UtilityNet-- high performance distributed intelligent computing network

UtilityNetaggregates scattered AI physical computing resources into a huge computing pool, and performs computing tasks in response to different development requirements. For developers, operating a distributed AI cluster resource can be as simple as using a computer.

The tasks include graphics processing, machine learning, data analysis and mining, and the latest deep learning technology. Among them, the deep learning module provides a variety of performance optimization schemes from data, algorithms and models, provides fine-grained scheduling and deployment, and accurately supports the exclusive sharing of specific AI computing resources such as TPU, GPU, CPU, memory, etc. to a single node, so that developers can focus on their own professional fields and develop efficiently.

Objectives in the first stage of development tests:

Support x86 cluster architecture, multi-node computing and networking

Support the allocation, recovery and task scheduling of computing resources of 100,000-level clients

Common deep learning and reasoning frameworks can be applied at the user end

Support the fine allocation of single TPU resources

NVMe+RDMA High-bandwidth low-latency mixed flash storage

AI computing cost 2-4 times lower than that of the commercial price

At the software resource level, UtilityNet will integrate common deep learning frameworks such as Caffe, Darknet, MXNet, ONNX, PyTorch, PaddlePaddle, TensorFlow, and common deep learning data sets such as Mnist, MS-COCO, ImageNet, etc., so as to provide developers with software and hardware solutions.