HPE, Nvidia partner on scalable hardware platform for AI development


Hewlett Packard Enterprise Co. (HPE) has entered the Nvidia Corp. alliance, unveiling a range of co-developed systems optimized for artificial intelligence (AI) during its HPE Discover conference in Las Vegas. These systems are designed to enhance productivity for AI engineers, data scientists, and operations teams by up to 90%.

HPE Private Cloud AI integrates Nvidia AI computing, networking, and software with HPE’s AI-specific storage and computing products, available through the GreenLake cloud platform. GreenLake offers cloud-like capabilities across various IT infrastructures with flexible pay-as-you-go pricing.

This comprehensive offering includes the new OpsRamp AI copilot for IT operations and is available in four configurations to support diverse AI workloads. It facilitates tasks such as inferencing, fine-tuning, and automated generation using proprietary data, ensuring data privacy, security, transparency, and governance controls alongside ITOps and AIOps capabilities.

At the core of the AI and data software stack is Nvidia AI Enterprise software platform and inference microservices, designed for rapid development and deployment of generative AI applications. HPE AI Essentials software complements this with curated AI and data tools, offering enterprise support and features for ensuring compliance, transparency, and reproducibility in AI pipelines.

Fidelma Russo, HPE’s chief technology officer, emphasized the simplicity of deployment: “It’s ready to boot up. You plug it in, connect to the cloud, and you’re up and running three clicks later.”

Neil MacDonald, executive vice president and general manager of HPE’s high-performance computing and AI business, highlighted the offering as an on-premises alternative to AI model training and inference in the cloud. He noted the cost-effectiveness of on-premises inferencing, citing it as four to five times cheaper while maintaining data privacy.

The infrastructure stack includes Nvidia Spectrum-X Ethernet networking, HPE GreenLake File Storage, and HPE ProLiant servers supporting Nvidia L40S, Nvidia H100 NVL Tensor Core GPUs, and the Nvidia GH200 NVL2 rack-scale platform.

Available in configurations ranging from small to extra-large, each system is modular and expandable to meet varying processing needs:

  • Small configuration: 428 Nvidia L40S GPUs, up to 248 terabytes storage, and 100-gigabit Ethernet networking in an 8-kilowatt rack.
  • Midsize configuration: Up to 16 L40S GPUs, 390 terabytes storage, and 200-gigabit Ethernet networking in a 17.7-kilowatt rack.
  • Large model: Up to 32 Nvidia H100 NVL GPUs, 1.1 petabytes storage, and 400-gigabit Ethernet networking in a twin rack configuration at 25 kilowatts each.
  • Extra-large model: Up to 24 GH200 NVL2 GPUs, 1.1 petabytes storage, and 800-gigabit Ethernet networking in a twin 25-kilowatt rack configuration.

HPE plans to support Nvidia’s GB200 NVL72/NVL2 and new architectures like Nvidia Blackwell, Rubin, and Vera. GreenLake provides management and observability services across hybrid environments, while OpsRamp offers visibility into Nvidia accelerated computing stack, including NIM and AI software, Tensor Core GPUs, AI clusters, and Quantum InfiniBand and Spectrum Ethernet switches.

HPE and Nvidia, alongside service providers Deloitte LLP, HCL Technologies Ltd., Infosys Ltd., Tata Consultancy Service Ltd., and Wipro Ltd., will market and sell these server packages.

In a separate initiative, HPE collaborates with engineering firm Danfoss A/S to develop an off-the-shelf heat recovery module aimed at reducing AI model development energy costs. This solution integrates HPE’s Modular Data Center with Danfoss’ heat reuse modules and oil-free compressors, enhancing data center cooling efficiency and energy consumption by up to 30%.

HPE’s initiatives underscore its commitment to advancing AI capabilities and sustainability in IT infrastructures, addressing diverse enterprise needs with innovative, scalable solutions.

Source: siliconangle.com