Setup and configuration of high-performance computing (HPC) systems can be a substantial obstacle that needs experienced IT pros to establish the software application stack, for example, and optimize it for maximum efficiency– it isn’t like constructing a PC with parts bought off NewEgg.GigaIO, which focuses on facilities for AI and technical computing, is aiming to simplify the job. The vendor recently revealed a self-contained, single-node system with 32 set up GPUs in package to provide simplified implementation of AI and supercomputing resources.Up to now, the only way to harness 32 GPUs would need four servers with eight
GPUs apiece. There would be latency to contend with, as the servers communicate over networking procedures, and all that hardware would take in floor space.What makes GigaIO’s device– called SuperNODE– notable is that it uses a choice of GPUs: up to 32 AMD Instinct MI210 GPUs or 24 NVIDIA A100s, plus as much as 1PB storage to a single off-the-shelf server. The MI210 is an action down in efficiency from the top-of-the-line MI250 card( at least in the meantime) that’s utilized in the Frontier exaFLOP supercomputer. It has a couple of less cores and less memory however is still based upon AMD’s Radeon GPU technology. “AMD works together with start-up innovators like GigaIO in order to bring distinct services to the progressing work needs of AI and HPC,”said Andrew Dieckmann, corporate vice president and general supervisor of the information center and sped up processing group at AMD, in a statement.” The SuperNODE system developed by GigaIO and powered by AMD Instinct accelerators uses engaging TCO for both conventional HPC and generative AI work.”SuperNODE is constructed on GigaIO’s FabreX customized material technology, a memory-centric fabric that minimizes latency from system memory of one server communicating with other servers in the system to simply 200ns. This allows the FabreX Gen4 implementation to scale approximately 512Gbits/sec bandwidth. FabreX can link a range of resources, consisting of accelerators such as GPUs, DPUs, TPUs, FPGAs and SoCs; storage gadgets, such as NVMe, PCIe native storage; and other I/O resources linked to compute nodes
. Generally, anything that uses a PCI Express bus can be connected to FabreX for direct device-to-device communication throughout the same fabric.SuperNODE has 3 moduses operandi: monster mode, for applications that take advantage of numerous or all GPUs; freestyle mode, where every user gets their own GPU to use for processing functions; and swarm mode, where applications work on multiple servers.
SuperNODE can run existing applications composed on popular AI structures such as PyTorch and TensorFlow without requiring adjustment. It utilizes Nvidia’s Bright Cluster Supervisor Data Science software application to handle and configure the environment and manage scheduling as well as container
management.SuperNODE is offered now from GigaIO. Copyright © 2023 IDG Communications, Inc. Source
