Silicon Motion Technology is strengthening its presence in the enterprise storage sector with SSD host controllers, targeting prominent large data clients in North America and edge AI customers in Taiwan.
The rapid expansion of global AI server construction has fueled an increase in data storage demand. While NAND original manufacturers actively deploy 2Tb QLC NAND, samples are expected to be available between the second half of 2024 and the first quarter of 2025.
Silicon Motion has taken a two-pronged strategy to target North American large-scale data center clients, as well as Taiwan-based edge AI customers, with volume production slated to begin in the second half of 2026, according to the Taiwan-based NAND device controller specialist.
AI servers drive demand for high-capacity SSDs
The key difference between AI servers and traditional servers is the substantial rise in data consumption, according to Silicon Motion SVP Alex Chou. AI servers have two to four times the storage capacity of conventional servers.
During the training process, the AI server must be operational at all times for a minimum of four months, Chou indicated. To safeguard data with substantial volumes, temporary storage requiring 14TB of capacity per unit is necessary.
The importance of high-capacity SSDs is increasing as AI servers become more prevalent, Chou said. Downstream clients are seeking to upgrade to 64TB or even 128TB SSDs.
The key enabler is that chipmakers' QLC NAND flash memory now has a single die capacity of 2Tb, which meets the specifications for ultra-large-capacity SSDs, Chou continued.
QLC technology enables cost-effective high-capacity storage
QLC offers superior storage density compared to TLC, allowing for a greater number of storage units to be extracted from the wafer. Consequently, QLC SSDs are more cost-effective per unit of storage capacity than TLC SSDs, meeting the substantial storage demands of AI servers while also decreasing power consumption, thereby serving as a replacement for traditional HDDs.
Nevertheless, market sources indicate that the number of chipmakers capable of producing 2Tb NAND chips remains relatively low. Kioxia and Micron Technology were among the first to manufacture 2TB NAND devices. Western Digital is expected to release 2Tb chip prototypes in the first half of 2025, while its Korean competitors are anticipated to release them in succession.
According to Chou, AI-driven demand for high-capacity SSDs necessitates advanced NAND controllers. Currently, the industry favors 16TB SSDs. However, in scenarios requiring extremely high capacity, the Flexible Data Placement (FDP) function becomes more important. Efficient operations require coordination with the SSD's data allocation and structuring features.
Silicon Motion has developed PerformaShape, a patented technology that enables SSD applications to attain configuration optimization, Chou said. The combination of QLC NAND, FDP, and PerformaShape proprietary technologies will enable AI SSDs to realize substantial advantages through the additive effect of these three key components.
Edge AI SSD market shows promise
Silicon Motion has also expressed optimism about the high demand for edge AI SSDs. AI servers are divided into two categories: processing and storage, based on market demand. In the processing sector, the company will target major US OEM clients or original NAND manufacturers.
Given that original NAND manufacturers have storage solutions, they are expected to work with Silicon Motion in the processor sector or provide these solutions to key US OEMs. Silicon Motion can use SDKs or partner with large clients that already have resources to create full firmware.
Furthermore, Silicon Motion has been collaborating with three Taiwanese AI SSD customers to develop a diverse array of customized designs for SSD end customers. These designs are subsequently delivered to AI server customers indirectly.
Dedicated storage IC solutions will be introduced in the future as a result of the increased demand for storage density in edge AI SSDs. The combination with QLC NAND memory is expected to result in cost savings.
Article translated by Jessie Shen
