Getty Images

WD unveils storage for AI framework alongside new drives

As the focus for enterprise AI spreads beyond compute, Western Digital introduces a new SSD and HDD. It also released an AI framework that outlines what storage should be used when.

Western Digital wants to support all stages of an AI model's lifecycle by identifying and building high-performance SSDs and high-capacity HDDs that are optimized for each stage of that cycle.

Western Digital introduced its AI Data Cycle, which outlines the six-stage continuous loop of an AI model, from raw archives to new content generation, and charts a storage type for each stage. Along with the framework, the vendor is releasing its first PCIe 5.0 SSD and expanding the range of another SSD it currently offers. It's also sampling a new, denser HDD in its portfolio of storage devices.

As generative AI has taken center stage in IT, the focus has been on GPUs and high-bandwidth memory (HBM), according to Joseph Unsworth, an analyst at Gartner. While GPUs and HBM are still in demand, compute has established an installed base and is generating data that now requires storage.

"In 2023, most customers over-rotated to compute [and] GPUs for AI capabilities, and storage was starved in a lot of cases to help pay for it," Unsworth said.

Plus, storage vendors have a better understanding of how to store generated data optimally, and those vendors -- including SSD and HDD makers -- are releasing products to meet those needs.

AI Data Cycle

Western Digital's AI Data Cycle is a framework designed to help users understand the role storage types can play in AI. The framework maps six stages of an AI model: raw data archives and content storage; data preparation and ingestion; AI model training; interface and prompting; AI inference engine; and new content generation.

For the first and sixth stages, Western Digital recommends dense low-cost storage, HDDs; for the data preparation and ingestion stage, it recommends high-speed flash; and for the AI model training, inference prompting and inference engine, it recommends fast, dense flash.

The expectation was that the uptick in AI data generation would eventually translate into companies needing more storage, according to Jeff Janukowicz, an analyst at IDC. After a few years of low demand for both HDDs and NAND, including a large price decrease, storage saw an uplift in the first quarter of 2024.

"Companies are stepping back and saying, 'Data is important, it's growing and it needs to be stored. The more data we store, the better our models will be,'" he said.

As organizations store more data, they are also finding out they need to modernize their infrastructure to better support AI workloads, Janukowicz said. Major OEMs such as Dell have focused on putting together an AI-centric architecture for customers, but modernization is also filtering down to the component or drive vendors such as WD.

Flash performance when it matters

For training, Western Digital released the Ultrastar DC SN861, the vendor's first enterprise PCIe Gen 5.0 SSD, in both an E1.S and U.2 form factor. The SN861 SSD delivers high-performance storage without utilizing some of the enterprise features typically associated with this type of drive, according to the vendor.

The SN861 is an example of how SSDs can be optimized, Janukowicz said. In the past, vendors offered different technologies, such as single-level cell flash -- where one bit is stored per cell, for more speed -- or triple-level cell flash, where three bits are stored per cell, for density. But vendors can change how the drive is set up, with more or less endurance, features or differing levels of overprovisioning.

What we're seeing now is WD adjusting its roadmap to tailor those specifics, to optimize how data is being used along the AI pipeline.
Jeff JanukowiczAnalyst, IDC

"What we're seeing now is WD adjusting its roadmap to tailor those specifics, to optimize how data is being used along the AI pipeline," Janukowicz said.

Western Digital expanded its enterprise-class SSD portfolio to include a higher range of capacities for its Ultrastar DC SN655. This higher range is for use cases where the balance between performance and density is more critical than performance alone. The SN655, which can be used for model training, inference prompting and inference engine stages, is shipping now in 15.36 TB but will be available in 32 TB and 64 TB drives in the third quarter of 2024.

There is a need for flash storage at this stage, as well as a demand to consolidate the number of individual drives, Unsworth said. Only one other vendor, Solidigm, offers SSDs over 60 TB.

"Through consolidation, this helps minimize overall power consumption and hit sustainability targets, which are at odds with growing AI demands," he said.

HDDs still important to AI

Western Digital includes the Ultrastar HC680, a 28 TB HDD, in its AI Data Cycle. It's also currently sampling its new Ultrastar HC690 32 TB HDD. Both use shingled magnetic recording (SMR), where one written data track slightly overlays the next for greater density, and OptiNAND technology, flash storage embedded in the drive to increase performance and density.

A drawback to SMR technology is having to write data sequentially due to the nature of overlapping shingles. This may slow down the write speed, but the technology is needed to get to increasing capacities, Janukowicz said.

While flash storage is important for ingestion, training and even inferencing, most enterprise data is sitting on hard drives, he added. The more training data available to an AI model, the more accurate the model results, and hard drives are the most cost-effective way to store primary data.

"Additional ways to store data by finding higher capacity points are what's needed … and hard drives are a great complement to flash," Janukowicz said.

Adam Armstrong is a TechTarget Editorial news writer covering file and block storage hardware and private clouds. He previously worked at

Dig Deeper on Flash memory and storage

Disaster Recovery
Data Backup
Data Center
and ESG