AI’s thirst for power pushing enterprises into supercomputing

Enterprises embarking on large-scale artificial intelligence (AI) initiatives are already stepping into the world of supercomputing, whether they label it as such or not, according to Trish Damkroger, senior vice-president and general manager of high performance computing (HPC) and AI infrastructure solutions at Hewlett Packard Enterprise (HPE).

Speaking to Computer Weekly in a recent interview, Damkroger noted that the fundamental principles underpinning modern AI infrastructure – massive compute power, high-density configurations and scale-up architectures – are direct parallels to traditional supercomputing.

“It’s all aligned with supercomputing, whether you want to call it supercomputing or not,” she said. “It's about dense computing and scale up architecture,” she added, pointing to burgeoning power demands as a clear indicator and citing discussions with customers about building one-gigawatt datacentres, which is becoming the norm.

While the term supercomputing might conjure images of research institutions, Damkroger said some sectors are also leveraging HPC to run AI applications. She cited a quant trading fund that is looking at using supercomputers which are more cost-effective for high-density AI workloads that require direct liquid cooling.

In addition, South Korea’s SK Telecom is also using supercomputing to train large Korean language models based on OpenAI’s GPT-3. These models power AI services and applications across the telco’s mobile network. HPE provided an integrated, high-performance architecture to support large-scale training and deployment.

In Japan, Toyo Tires adopted HPE GreenLake with HPE’s Cray XD systems to speed up simulations for tire design. With up to three times faster performance, the company can now run complex, large-scale simulations in half the time, accelerating product development through HPC and AI.

Indeed, the growing adoption of AI has spearheaded interest in HPC systems across the Asia-Pacific (APAC) region. “Last year, our AI sales in APAC was second after North America, which is not normally the case,” Damkroger said. “There’s a lot of growth in the AI space in the region.”

To cater to diverse enterprise needs, HPE offers a flexible software strategy. This includes AI factories that let customers select open-source frameworks on top of HPE’s cluster management software, orchestrated by the Morpheus hybrid cloud management platform. For those looking for more turn-key capabilities, Damkroger said HPE’s Private Cloud AI is a curated offering that will let AI and IT teams experiment with and scale AI projects. “It's a one click easy button to get AI working,” she added.

Despite the advancements and growing adoption of AI, finding the truly transformative enterprise AI application that leverages HPC remains an ongoing quest. “If you look at AI specifically in enterprises, there are some good use cases, but I don’t think we found the most amazing ones yet,” Damkroger admitted.

While internal efficiencies, such as using large language models for service documentation, are valuable, “I don't know if we’ve found that killer app yet that it’s worth a lot of the cost overhead,” she said, noting that key challenges for broader enterprise HPC adoption for AI include the initial infrastructure investment, power requirements and a persistent talent shortage.

Addressing the common enterprise starting point of leveraging public cloud for HPC and AI workloads, Damkroger said that for sustained and intensive use, on-premises HPC deployments are more economical.

“We find that in the HPC space, if you are going to use it a lot with greater than 70% utilisation, it’s much more cost-effective to have it on-premises,” Damkroger said. However, she acknowledged the public cloud’s role for exploration and lower-demand scenarios, adding that the security of data is also a pivotal factor that tips the scales towards on-premises for sensitive HPC workloads.

Reflecting on HPE’s deep roots in HPC, including the Cray heritage which will mark 50 years since its first supercomputer this year, Damkroger said: “It’s fun to be in this business right now with liquid cooling being so prominent. We’re finally seeing all the work we've been doing for the last 50 years and we're able to take advantage of it.”