Samsung’s archrival wants to spread 100 special CPUs in every SSD controller ahead of PCIe 8.0, High Bandwidth Flash onslaught
- SK Hynix plans to distribute more than 100 processors across SSD controllers
- H-TPUs reportedly consume just 2.5% of conventional CPU core power
- Each H-TPU occupies only 4% of a general-purpose processor core’s area
SK Hynix is developing solid-state drive controllers with more than 100 compact processors to manage increasingly demanding data transfers.
The South Korean memory manufacturer says its Hynix Transaction Processing Unit (H-TPU) architecture reduces processor power requirements and chip area.
The technology is designed for enterprise storage products and future memory systems serving demanding artificial intelligence workloads.
Latest Videos FromTechRadar
Small processors take over repetitive SSD operations
Speaking at the recent International SoC Design Conference, SK Hynix vice president Koo In-je described limitations affecting conventional SSD controller designs.
“Currently, general-purpose CPUs inside SSD controllers are at their limits in terms of power consumption and area,” he said.
“Instead, SK Hynix has developed an H-TPU that can efficiently handle common input/output (I/O) transmission paths.”
Conventional processors handle routine operations alongside demanding tasks, including reading and writing data, updating metadata, and controlling buffers.
Faster storage interfaces increase these workloads, potentially requiring additional processor cores and increasing the controller’s physical footprint and electricity consumption.
H-TPUs separate repetitive operations from complicated processing, allowing numerous specialized cores to handle individual responsibilities.
SK Hynix says each processor uses just 4% of the chip area occupied by a general-purpose CPU core and consumes approximately 2.5% of its power.
The architecture uses a four-stage pipeline and sequential execution, which SK Hynix considers suitable for recurring storage operations.
Rather than activating every processor continuously, the controller enables the units required for each incoming request and places unnecessary processors into idle mode.
Conventional CPU cores remain responsible for demanding functions such as error recovery, exception management, and storage policy decisions.
This division allows specialized processors to handle routine work while the main controller retains responsibility for more complex operations.
SK Hynix plans H-TPUs for next-generation storage
SK Hynix intends to incorporate H-TPUs into upcoming enterprise SSD controllers and its Storage Next architecture.
The company also plans to use the technology with High Bandwidth Flash (HBF), a memory technology expected to enter the market next year.
Storage Next is designed to handle smaller data requests associated with newer AI models, including transactions involving 512 bytes or less.
That represents one-eighth of the conventional 4KB data unit, allowing storage systems to access smaller portions of information.
Future PCIe generations, including PCIe 6.0, 7.0 and 8.0, are expected to increase transfer speeds, placing additional demands on storage controllers.
Distributing repetitive processing across many small cores could help limit power consumption and physical expansion as those demands grow.
SK Hynix also intends to extend the technology to its AIN storage portfolio, which includes the AIN P, AIN B and AIN D product families.
The company has not disclosed when H-TPUs will reach individual commercial products or quantified their effects on overall SSD performance.
Via The Elec (originally in Japanese)
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds.
Share this article Join the conversation Follow us Add us as a preferred source on Google Newsletter Subscribe to our newsletter SK Hynix plans to distribute more than 100 processors across SSD controllers H-TPUs reportedly consume just 2.5% of conventional CPU core power Each H-TPU occupies only 4% of a…
