4U 8x A100 GPU Server for AI Training and Inference

Versatile 8GPU Server for AI Workloads

Ranking first in single server performance in MLPerf 0.7, it supports 8 fully interconnected NVLink 3.0 NVIDIA A100 Tensor Core GPUs, and 2 AMD EPYC™ Rome processors enabling PCIe4.0 in a 4U space, NF5488A5 delivers outstanding AI performance of 5 petaFLOPS and ultra-high data throughput, making it ideal choice for AI applications like image, video and voice processing, financial analysis and virtual assistants.

Key Features

Record-Setting AI Performance

Best single server performance in the MLPerf v0.7 benchmark, setting 18 performance records and achieving 549,000 inferences per second in the ResNet50 inference task.

Learn more: Press Release | Blog Post | Official MLPerf v0.7 Results

Extreme Hardware Design

8x fully interconnected NVLink 3.0 NVIDIA A100 GPUs, 2x AMD EPYC™ Rome processors enabling PCIe4.0 in 4U and delivering 5 petaFLOPS and high data throughput.

Accelerating AI Training

Supports the latest A100 80G GPU with HBM2e, doubling the memory capacity of its predecessor and a 30% increase in GPU memory bandwidth in a single node, delivering 15% performance increase in AI model training tasks with tens of billions of parameters.

Peek under the hood of Inspur’s NF5488A5 to see how this 4U 8x A100 GPU server delivers leading-edge, record-breaking 5 petaFLOPS performance for AI training and inference applications.

Take a dive into the NF5488A5 with Serve The Home, who share a detailed look at features like the 8x A100 GPU assembly, optimized hardware design, and MLperf-chart-topping performance.

The Experts Weigh In on NF5488A5

Hardware industry specialists Serve The Home gives the NF5488A5 a 9.7 out of 10 in this comprehensive review, where they break down the server’s features like hardware design and performance.

Read the Review »

Model	NF5488A5
GPU	1* HGX A100 8GPU
Processor	2* AMD EPYC™ Rome Processor (Up to 225W TDP)
Performance	5 petaFLOPS
Memory	32* DDR4 2933MT/s RDIMMs//LRDIMM, up to 4 TB
Storage	8* 2.5” SAS/SATA (or 4NVMe+4 SATA/SAS), 4* NVMe M.2, 2* SATA M.2
PCIe	4* built-in low profile PCIe x16
Front I/O	2* USB 3.0 port, 1* VGA port, 1* RJ45 management port
Ethernet Card	10G Ethernet Optical Interface
Management	Built-in Aspeed2500 BMC module, support IPMI, SOL, KVM Over IP, VM, etc.
RAID	Support 3508 Raid Card (NVMe Raid with Raid Key)
OS	Red Hat Enterprise 7.8 64bit, CentOS 7.8; Ubuntu 18.04
Cooling	N+1 Redundant hot swap fans
PSU	3+1 Redundant PSUs
Chassis	4U, 448mm W x 175.5mm H x 850mm D

Interested to learn more?

Inquire Now

	Record-Setting AI Performance Best single server performance in the MLPerf v0.7 benchmark, setting 18 performance records and achieving 549,000 inferences per second in the ResNet50 inference task. Learn more: Press Release \| Blog Post \| Official MLPerf v0.7 Results

	Extreme Hardware Design 8x fully interconnected NVLink 3.0 NVIDIA A100 GPUs, 2x AMD EPYC™ Rome processors enabling PCIe4.0 in 4U and delivering 5 petaFLOPS and high data throughput.

	Accelerating AI Training Supports the latest A100 80G GPU with HBM2e, doubling the memory capacity of its predecessor and a 30% increase in GPU memory bandwidth in a single node, delivering 15% performance increase in AI model training tasks with tens of billions of parameters.