HPE Discover, Las Vegas – HPE (NYSE: HPE) jumped into the artificial intelligence (AI) cloud market today, announcing it will now offer large language models (LLMs) for enterprises and startups to train and deploy AI through its supercomputing cloud service.
John Abbott, an infrastructure analyst at 451 Research, said HPE’s model could offer the company an advantage despite competition from hyperscalers like Microsoft and Google.
"To run its new cloud service – HPE Greenlake for Large Language Models – HPE has built a more stripped down, more modular system that’s easier to upgrade," Abbott told Silverlinings. That's a good thing since "the current pace of technology in this sector is blistering."
The new on-demand service — which will be powered by nearly 100% renewable energy — will be delivered in partnership with German AI startup Aleph Alpha. The latter will supply customer use cases with Luminous, a “field-proven and ready-to-use LLM,” HPE said in a release.
“By using HPE’s supercomputers and AI software, we efficiently and quickly trained Luminous for critical businesses such as banks, hospitals and law firms to use as a digital assistant to speed up decision-making and save time and resources,” Aleph Alpha’s CEO and founder Jonas Andrulis stated in the release. Luminous will be offered in English, French, German, Italian and Spanish.
HPE’s EVP and GM for HPC & AI Justin Hotard said during a press briefing, “We see this as a complementary offering [to other AI products]. We don't see this as competitive” since the roadmap to wider AI adoption will demand intimate collaboration. The company plans to continue looking for best-of-breed models and partners to work with. Additionally, HPE plans to “integrate with the public cloud” via APIs to help enterprises with data residing on public clouds integrate with HPE’s services.
Looking under the hood, the service will initially run on Cray XD supercomputers — HPE acquired Cray for $1.3 billion in 2019 — and the LLMs will run on NVIDIA's H100 GPUs.
“HPE and Cray have offered HPC services in the cloud before,” Abbott told Silverlinings. “They set up a service with Microsoft Azure in 2019 (Cray ClusterStore in Azure)...This time there’s a more urgent demand for high-end resources to train generative AI.”
HPE’s supercomputing platform uses HPE’s software suite, the HPE Cray Programming Environment, to optimize the high performance compute and AI applicaitons with tools for developing, porting, debugging and tuning code, the release stated. Still, Hotard noted there will be future announcements regarding hardware based on specific use cases.
AI ‘inflection point’
AI is at an “inflection point,” according to Hotard, and to realize the benefits it promises to deliver, it will be essential that AI projects have key foundational capabilities to support “their journey from proof of concept to scaled deployment.”
Hotard divided those capabilities into four categories: access to scalable computing, since training and tuning large AI models require specialized resources; acquired talent for installation and management of these resources; security and trust leveraging data to train the model responsibly; and an optimized software layer to support the workloads at scale.
These pillars create a structured environment in which “AI models can train more efficiently over a single cluster of nodes all operating as one single computer,” Hotard explained.
HPE is initially targeting enterprise customers and potential startups looking to accelerate their own services, but it also expects public sector customers. The pricing is currently undisclosed, only sold direct, but Hotard said HPE will continue to evaluate other routes to market and drive price based on size and use of resources. Future partners following Aleph Alpha will be able to make their models available on their own pricing terms.
Looking at the broader picture, 451 Research's Abbott said “There will be competition from the cloud hyperscalers as they build out their own massive GPU farms,” but added there “are some aspects of the cloud model that might give HPE an advantage — it won’t, for instance, add punitive data egress fees, and its infrastructure can handle single, large scale supercomputer and AI jobs that must run uninterrupted for 24 hours and more — something public clouds were never designed for.”
This article was amended on Weds., June 21, 2023, at 10:00 AM ET.