Large language models (LLMs) are set to power modern enterprise. Still, developing them can be resource-intensive, to say the very least.

The solution to effectively train bigger and more accurate models: Supercomputing, according to Justin Hotard, executive VP and GM for the HPC and AI business group at Hewlett Packard Enterprise (HPE).

“Supercomputing provides massive performance and an optimized architecture that runs as a single computer,” he said.

That said, supercomputing can be out of reach of many enterprises. To help bridge this gap, HPE announced today at HPE Discover 2023 its new platform HPE GreenLake for Large Language Models (LLMs).

The on-demand, multi-tenant supercomputing cloud service will allow any enterprise to leverage supercomputing and privately train, tune and deploy “massive AI” without having to invest in their own supercomputer, according to Hotard.

The platform is the first in a series of industry and domain-specific AI applications planned by HPE; future support will come for climate modeling, healthcare and life sciences, financial services, manufacturing and transportation. Furthermore, HPE GreenLake for LLMs will run on supercomputers and AI software that are powered by “nearly 100%” renewable energy, according to the company.

“We're committed to making our supercomputers accessible for everyone to capitalize on AI,” Hotard said in a pre-briefing.

Scaling AI workloads for large-scale production can be problematic for enterprises, Hotard pointed out. This is due to the increased need for on-demand capability computing and the lack of human, financial and technical resources. Also, security and multi-tenancy have different demands, and traditional public clouds are not optimized for strong scaling jobs.

Supercomputing allows single large-scale AI and high performance computing (HPC) jobs to run on hundreds or thousands of CPUs or GPUs at once, which is more effective, reliable and efficient for training AI and creating more accurate models, he said.

This is “very, very different from general purpose cloud offerings that run multiple jobs in parallel on a single instance,” he said.

But supercomputers are costly and complex to adopt and manage and also require unique data center infrastructure, skills and capabilities when it comes to power and cooling, Hotard pointed out. Also, they have generally not been available on demand in a consumption model.

On the other hand, if an enterprise chooses to purchase high performance computing and AI from hyperscalers, this comes at the cost of  unnecessary data egress and limited control of sensitive models and data. Also, cloud is really not designed for single, large-scale supercomputer and AI jobs, Hotard said.

“Access to scalable computing, training and retraining of large AI models requires specialized computing resources — what we call capability computing — that is not readily available in most existing data centers or in the public cloud,” he said.

HPE GreenLake for LLMs runs a single AI workload at full computing capacity. It includes access to Luminous, a pre-trained LLM from Aleph Alpha, which supports several languages including English, French, German, Italian and Spanish.

“Users can upload their own data and securely protect, train and tune their own customized model solely for their use,” said Hotard.

The new LLM tool will run on HPE Cray XD supercomputers, and customers will have access to HPE’s open-source tools. Specifically, the HPE machine learning development environment helps rapidly train and scale models — such as generative AI — with intelligent hyper parameter optimization. And, HPE machine learning data management software provides visibility by integrating, tracking and auditing data.

Finally, HPE will support a software model library that will include both open source models or proprietary third-party models (such as the Luminous LLM).

“This integrated stack means that enterprises can bring their requirements and their data to quickly and train to train and deploy large language models,” said Hotard.

But, he pointed out that HPE isn’t fully dismissing the public cloud: In fact, the company intends to eventually integrate with leading cloud providers. As he put it, data is critical to training and tuning models and needs to be retrieved from wherever it resides (on-premises or in the cloud).

“We see this as a complementary offering,” he said. “We don't see this as competitive; we believe that there will be opportunities for partnership.”

Just as importantly, he emphasized HPE’s commitment to sustainability. The company’s goal is to provide 100% carbon neutral offerings, notably through renewable power in the way of liquid cooling. HPE GreenLake for LLMs will run in colocation facilities such as with QScale in North America.

Ultimately, HPE president and CEO Antonio Neri said in a statement: “We have reached a generational market shift in AI that will be as transformational as the web, mobile, and cloud.”

HPE’s goal is to “make AI, once the domain of well-funded government labs and the global cloud giants, accessible to all.”

Also announced today are the following:

– OpsRamp is now available as a SaaS offering on the HPE GreenLake platform. The AIOps-powered IT operations management tool provides full stack observability across multi-vendor, multi-cloud infrastructure and applications.

– New HPE GreenLake Software-as-a-Service (SaaS) offerings for backup and machine learning, and extended Network-as-a-service (NaaS) portfolio.

–Expansion of the HPE GreenLake private cloud portfolio to address edge use cases and provide container support for Red Hat.

– New HPE GreenLake for Private Cloud Business Edition, which allows customers to spin up virtual machines (VMs) across hybrid clouds on demand.