HPE targets enterprises with Nvidia-powered platform for tuning AI

'We feel like enterprises are either going to become AI powered, or they're going to become obsolete'

HPE Discover EMEA HPE - like many tech companies - is betting big on AI in the hope customers will splash enteprise cash on training or fine tuning models and other areas of interest, rather than risk falling behind their peers.

At the annual Discover event, this year held on Barcelona, HPE lifted the lid on a second generative AI platform co-engineered with GPU maker Nvidia just weeks after the first, and also confirmed that the HPE Machine Learning Development Environment is now available as a managed service on public clouds, starting with AWS and Google Cloud.

HPE's prosaically named "enterprise computing solution for generative AI" is a pre-configured platform comprising a mix of HPE hardware and software plus Nvidia's GPUs, networking and AI software.

The pair recently announced a supercomputing system for AI that follows a similar theme, but that is built to train generative AI models. This latest platform is instead a more modest one intended for enterprise customers to tune existing models to their requirements and then operate them for inferencing work.

HPE describes the enterprise computing solution for generative AI as a rack-scale architecture, meaning that it pulls together multiple components which effectively fill a datacenter rack, or most of one. The compute in this case is provided by 16 ProLiant DL380a Gen11 servers based on Intel Xeon Scalable processors, which can be configured with up to four Nvidia L40 GPUs.

Networking in this platform is provided by Nvidia's Spectrum-X switches and network adapters based on its BlueField-3 DPU chip.

This configuration is sized to allow it to fine-tune a 70 billion parameter version of Meta's Llama-2 model, according to HPE.

The software to run on this infrastructure is similar to that included with the supercomputing AI rig, namely the HPE Machine Learning Development Environment, and Nvidia's AI Enterprise suite and the NeMo framework for conversational AI.

HPE reiterated that AI calls for a new compute architecture, one that HPE hopes to supply, of course.

"AI requires a fundamentally different architecture, because the workload is fundamentally different than their classic transaction processing and web services workloads that have become so dominant in computing over the last couple of decades," said Evan Sparks, HPE's chief product officer for AI.

"We think that the next decade is going to require full stack thinking from hardware all the way up to the software layer, as organisations lean into the deployment of AI applications," he added.

However, for many enterprises, finding their way with AI means taking existing models and experimenting to see if they add value into their business processes.

"Many organizations are not going to be building their own foundational models, they're going to be taking a model that has been developed elsewhere, and they're going to deploy it into their business to transform their business processes," said Neil MacDonald, EVP & general manager of HPE's Compute business.

"One of the challenges is building and deploying infrastructure that enables experimentation and fine tuning and then deployment of these models," he claimed. "We feel like enterprises are either going to become AI powered, or they're going to become obsolete."

What's the damage?

Like the supercomputer for AI, HPE has not yet detailed how much this is going to cost customers, but the enterprise computing solution for generative AI is set to be available to order some time in the first quarter of 2024.

HPE also said its Machine Learning Development Environment is now available as a managed service on public clouds. This is a platform to train up generative AI models, and largely based on technology HPE gained from its purchase of Determined AI in 2021.

"This is a fully managed service, fully managed by HPE, meaning that the end users don't have to worry about managing the infrastructure at all in their cloud or cloud accounts," claimed Sparks. It is available first on "popular platforms" like AWS and Google, he added.

The fully managed model is intended to reduce the complexity of AI/ML model training and thus speed the development process, HPE said.

HPE is also beefing up its Greenlake for File Storage offering to keep up with the data demands of AI workloads.

"Starting in Q2, we're significantly growing that for customers who want to scale up to you know, somewhere in the realm 250 petabytes of data," said Patrick Osborne, SVP and GM of Cloud and Data Infrastructure.

Also coming is support for Nvidia's Quantum-2 networking to allow customers to plug into InfiniBand fabrics for higher throughput, Osborne said.

HPE said the products will be available via the channel and via its subscription-based Greenlake finance model.

"Not everyone wants to spend a ton of CapEx right up front to be able to jump into this opportunity, so we can provide these through very flexible consumption models for our customers, and provide them flexibility both on the technology side as well as financially on the economic side," said Osborne. ®

More about


Send us news

Other stories you might like