Servers & Systems: The Right Compute

Intelligence for interconnect: AI workloads demand the scale, speed and versatility of Intel® OPA

When it comes to interconnect fabric, AI, machine learning and deep learning workloads all demand the scale, speed and versatility of Intel® Omni-Path Architecture (OPA) combined with HPE’s complete AI portfolio.

HPE HPC_Intel_AI_blog.jpgOnce merely an academic or pure research pursuit, artificial intelligence (AI) is at the forefront of enterprise IT. AI use cases being developed and implemented currently include fraud detection, cyberattack prevention, voice and image recognition and preventive maintenance.

Entire industries, too, such as oil and gas, are reorienting their business models to accommodate how AI and machine learning are changing their game.[1]

Yet making AI a realistic part of your company’s IT strategy requires a comprehensive, purpose-built portfolio for AI.

HPE and Intel represent a powerful partnership that is making AI work from the edge to the cloud, now and into the future. HPE offers a comprehensive portfolio of compute, storage, networking, software and services, helping customers with AI environments.

Key takeaways on HPE and Intel OPA “better together” AI solutions

-Overlooked question in architecting an AI cluster: What will be your interconnect fabric?

-Intel® Omni-Path Architecture (OPA) represents an important choice for scalable, energy-efficient, and robust AI implementations

-HPE servers and storage interface seamlessly with OPA together offering comprehensive, purpose-built portfolio for AI success

However, there’s an elemental hardware concern across all of these infrastructure components. Indeed, not enough businesses appreciate the value of another factor in securing their AI system’s success:

What fabric will you use to interconnect your AI cluster?

This question is important because when AI systems grow and scale up, as they often do, the fabric that stitches that system together must be able to grow seamlessly too—maintaining its speed, security, agility, versatility and robustness throughout.

Fundamentally, Intel OPA is an interconnect developed originally for high-performance computing (HPC) clusters, whose efficiency and speed in this domain improve scalability and increase density, while reducing latency, cost and power on the frontiers of AI as well.

By contrast, Ethernet—which might operate in a small cluster or AI proof of concept (POC)—was never designed to handle dozens or hundreds of nodes that AI, machine learning (ML) and deep learning (DL) training systems have been known to scale up to.

Moreover, clusters built with OPA can occupy a versatile niche in which they run HPC workloads during the day and compute-intensive deep learning training workloads at night.

Learn more about accelerating AI with high-performance fabrics by registering to watch this webinar.

HPE and Intel OPA: better together

If OPA is the interconnect fabric with which to architect a cutting-edge AI, ML or DL solution, the ideal cluster and storage that it interconnects comes from the same company who helped develop and enhance customized OPA switches and adapters in tandem with Intel.

HPE offers a comprehensive AI portfolio of compute, storage, networking, racks, software and services, helping customers with AI, ML and DL initiatives. HPE Apollo Systems, HPE SGI 8600 and HPE ProLiant Servers interface with OPA seamlessly, providing tight integration, density, low power, lower cost (compared to leading alternative fabric) and performance.

For instance, HPE Apollo 2000 Gen10 System can be used for both training models in the data center and as an inference engine on the edge. The HPE ProLiant XL190r Gen10 Server goes into the HPE Apollo 2000 chassis. This flexible server has two processors and additional PCIe slots in multiple configurations, providing support for additional expansion cards and support for two accelerators per server.

The HPE Apollo 6500 Gen10 System is an ideal HPC and deep learning platform providing superior performance-per-dollar for GPU intensive workloads with up to eight high performance GPUs per node delivering up to 125 TFlops of single precision compute.[2]

Then, for petaflop scale deep learning, the liquid-cooled HPE SGI 8600 delivers leading performance, density and efficiency and is used to train extremely, large deep learning models. The Tokyo Institute of Technology has been one of the most innovative teams of HPE SGI 8600 users, leveraging the OPA interconnect for a range of AI and HPC workloads.

 “Through [their] partnership with SGI, and now HPE, the Tokyo Institute of Technology has worked successfully to deliver a converged world-leading HPC and deep learning platform that can address our requirements and those of our nation.”  – Satoshi Matsuoka, TiTech professor and leader of the team that operates the TSUBAME 3.0 HPE SGI 8600 cluster

To discover how OPA and HPE can offer “better together” AI solutions to accelerate your business today, watch the OPA AI webinar or contact an Intel or HPE sales representative today.

The HPE and OPA discussion continues in this blog: HPE Systems with Intel® Omni-Path: Architected for value and accessible high-performance computing.

bill.jpgMeet Server Experts blogger Bill Seidle, Group Manager of High Performance Computing & Artificial Intelligence Solutions.  With over 22 years of marketing acumen in the highly competitive field of IT, Bill leads the HPE HPC & AI marketing team, focusing on delivering the best-in-class portfolio to help customers unleash their data insights faster. 

Server Experts
Hewlett Packard Enterprise

twitter.gif @HPE_HPC


[1] NVIDIA testing; Tesla V100 provides 15.7 TFlops single precision performance per GPU x 8 GPUs = 125 TFlops; published as of March 2018.

[2] “How Artificial Intelligence is Taking Over Oil and Gas”; published Aug. 10, 2018.




0 Kudos
About the Author


Our team of Hewlett Packard Enterprise server experts helps you to dive deep into relevant infrastructure topics.