DDN launches AI inference storage platform with Supermicro and Solidigm

The announcement came during the Supermicro Open Storage Summit interview series, where executives from the three companies discussed the collaboration and emerging trends in enterprise AI. Andrew Murphy, senior director of product management at DDN, said customers often choose best-of-breed components, but this can lead to complexity. "The question is how can we help to offload that complexity from our customers so they can focus on what's most important, which is taking their data that they've been generating over the years ... [and] add context so that it can be ready for their AI factories," he said.

Michael Ang, director of storage solutions at Supermicro, noted that customers traditionally buy compute, storage and networking separately and assemble them themselves, which takes too long. A unified solution enables organizations to start using and monetizing their data immediately, he argued. "The number one challenge is [customers] still have the mindset of having to buy individual — compute, storage, networking — and put it together, but that takes a long time," Ang said. "In order to help them be successful in this area, the customer should focus on how do they get the right solutions with the right platforms such that ... they can start utilizing it and start doing the AI workloads."

Pompey Nagra, products and ecosystems lead at Solidigm, said the HyperPOD can employ different types of flash memory to handle various AI tasks. Because solid-state drives can reuse AI workloads without recomputation, customers get more utilization out of their GPUs. "We see the platforms change from being compute bound to storage bound, where LLMs need more data to more storage to enable the [key-value] cache data," Nagra said. "The SSD storage becomes a great value as it's capacity bound versus the [high-bandwidth memory] or primary memory, which requires greater amounts of recomputation for the same workload."

The platform is built on Nvidia's secure reference design and can be deployed on premises, giving organizations greater control over security, governance and data residency. It also supports multi-tenancy, allowing multiple teams to share infrastructure while keeping their environments separate. Murphy said that GPUs are often idle while waiting for data, and DDN's focus is to provide a fast path to fully utilize GPUs.

The executives also highlighted growing demand for private and sovereign AI. Nagra said organizations want control of their data, IP and costs, and the platform allows them to operate within their governance, whether on premises or in hybrid environments. HyperPOD can scale from one rack to many as needed, reducing cost exposure and letting customers expand without committing to large numbers of expensive GPUs upfront. "We are able to design this complete AI data platform solution that allows the customer to expand from small, medium, large," Ang said. "Meaning they can continue to run the business when they see there's a need for additional GPUs and storage utilizations. They can just pop in another rack next to it, connect the network cables, power it on, then continue online and serve customers."