SK hynix outlined the H-TPU architecture at the International SoC Design Conference (ISOCC 2026) earlier this week. Vice President of system-on-chip design Koo In-jae said that SSD controller CPUs have reached their limits on power and area. PCIe bandwidth doubles every generation, he noted, and controller CPU clocks and control capability cannot keep pace.
The H-TPU is a small core with a four-stage, in-order pipeline. It is not superscalar and does not use out-of-order execution. It is designed for specific tasks: read and write command processing, metadata management, and buffer control. Each H-TPU draws about 2.5% of the power of a general-purpose CPU core and occupies roughly 4% of its area.
SK hynix broke more than 100 detailed storage control functions across “dozens of small TPU clusters,” according to reports from ZDNet Korea. The total number of H-TPUs exceeds 100, though the number of cores per cluster has not been disclosed. The main Arm CPU still handles error recovery, exception handling, and storage policy management.
The design is a response to the scaling challenge evident in competitor products. Marvell’s PCIe 6.0 Bravera SC6 controller uses 12 Cortex-R82 cores in two clusters plus three Cortex-M7 cores and a secure Cortex-M3, up from 10 Cortex-R8 and Cortex-M7 cores in its PCIe 5.0 predecessor.
SK hynix plans to apply H-TPU across its NAND products, including enterprise SSD controllers, where demand for high IOPS is most pressing. The cores come with a dedicated compiler and development tools for specialized code.