The claims are large, but they come from a conference talk. They are not from a shipping drive.
What SK hynix actually disclosed
Vice president Koo In-je told the International SoC Design Conference (ISOCC) 2026 in Incheon on October 12 that general-purpose CPUs in existing SSD controllers are reaching their limits in power consumption and chip area. The Elec reports that SK hynix has integrated more than 100 H-TPUs into its SSD controller architecture.
H-TPU stands for Hynix Transaction Processing Unit. The name invites confusion. It is not an AI NPU. It is a very small programmable processor for simple, repetitive tasks inside an SSD controller. Tom's Hardware likewise says it should not be mistaken for a tensor processor.
The design points reported so far:
- Pipeline: Each H-TPU has a four-stage pipeline and sequential execution, which suits repetitive workloads.
- Workloads: The cores take over read and write command handling, metadata management and buffer control from the controller's general-purpose CPUs.
- What stays on the main CPU: The general-purpose CPU is reserved for complex tasks such as error recovery, exception handling and storage-policy management.
- Power gating: When an I/O request arrives, only the H-TPUs assigned to the needed functions wake up. The rest stay idle to save power.
- Tooling: The Elec says SK hynix also built software tools, including a dedicated compiler, so engineers can write code for the cores.
This is offloading, not replacement. The Arm CPU stays in the controller. It just stops doing the grunt work.
The 2.5% and 4% figures
SK hynix says a single H-TPU uses about 2.5% of the power of a standard general-purpose CPU core. The Elec adds that each one occupies about 4% of the chip area.
Those numbers need careful reading:
- They compare one H-TPU with one CPU core. They do not describe a whole controller or a whole drive.
- Tom's Hardware notes that neither report names the CPU core behind the 2.5% and 4% figures.
- Hardware Busters points out that a controller with a hundred H-TPUs plus a main CPU still has to be measured against one with a few Arm cores, and that is the comparison that matters.
- Guru3D stresses that the figure comes from the company's presentation and is not an independent measurement from a shipping SSD.
So don't read this as a 97.5% cut in SSD power. A hundred cores at 2.5% each is not zero. Whether the total comes out ahead depends on how many are awake at once, and on what the workload looks like.
Why controllers are under pressure
Tom's Hardware reports that Koo's talk, titled "From SSD Controllers to AI Memory SoCs," made the case that PCIe bandwidth doubles each generation. It adds that controller CPU clocks and control capability can't keep pace.
The industry has mostly answered by adding cores. Tom's Hardware gives an example: Marvell's PCIe 6.0 Bravera SC6 uses 12 Cortex-R82 cores in two six-core clusters. It also has three Cortex-M7 cores and a secure Cortex-M3. Its PCIe 5.0 predecessor, the SC5, lists 10 Cortex-R8 and Cortex-M7 cores plus a Cortex-M3.
The Elec reports Koo's argument that Arm cores are overpowered for simple, repetitive SSD work. They are tuned for fast complex computation. They are also awkward to split into many small, parallel tasks.
Splitting the work is not new, either. Tom's Hardware says Phison already offloads repetitive firmware work. Its enterprise X1 controller used two Cortex-R5 cores plus up to 32 "micro cores." In that framing, H-TPU is the same idea at a much larger scale.
Where it is headed: Storage Next and HBF
The Elec reports that SK hynix plans to use H-TPUs first in next-generation enterprise SSD controllers. After that come:
- Storage Next: The Elec describes this as SK hynix's AI-oriented storage architecture. It handles data units down to 512 bytes, one-eighth of the conventional 4 KB unit. Tom's Hardware ties the name to Nvidia's naming for the joint AIN P project with SK hynix.
- High Bandwidth Flash (HBF): The Elec says HBF is scheduled for release next year. Tom's Hardware describes it as stacked NAND for AI inference. It aims for HBM-class bandwidth with far more capacity, up to 512 GB per stack against 64 GB for HBM4.
- The AIN family: TechRadar says SK hynix also intends to extend the technology to its AIN P, AIN B and AIN D product lines.
On HBF, Tom's Hardware says Sandisk and SK hynix released the formal specification in August through the Open Compute Project. That makes it an open standard rather than a proprietary interface. The H-TPU link to HBF, though, is a stated plan from the Elec's report, not a confirmed shipping design.
What is not established
The headline mentions PCIe 8.0. Here is what the evidence supports:
- TechRadar says future PCIe generations, including 6.0, 7.0 and 8.0, will push transfer speeds up and add to controller demands.
- Tom's Hardware says simple core scaling is suboptimal even at current PCIe speeds. It adds that PCIe 8.0 is headed for final ratification in 2028.
- No source reviewed says an H-TPU product has been validated for PCIe 8.0. No source names a PCIe generation for an H-TPU product, either.
Other gaps are just as plain:
- Guru3D reports that no enterprise controller using H-TPU has been announced, and no commercial release date has been provided.
- TechRadar notes that SK hynix has not quantified the effect on overall SSD performance.
- Tom's Hardware says H-TPU has yet to be proven in an actual enterprise SSD controller product, and that it knows of no date.
- Cluster sizes and scheduling overhead are undisclosed. Tom's Hardware says it has no details on how many cores each cluster holds.
Why Windows and PC readers should care
This is a data-center story first. The first target is enterprise SSDs, and HBF is aimed at AI inference systems. Nobody has promised H-TPU in a consumer NVMe drive.
Still, controller design shapes what ends up in PCs. The same pressure on power, heat and die area hits client drives as PCIe 5.0 SSDs run hot. Techniques proven in enterprise parts often trickle down. That is general industry experience, not something SK hynix has claimed here.
For admins, the useful questions to put to vendors once real products appear are:
- What are the whole-drive power and latency numbers, measured under mixed workloads?
- How does the drive behave under error recovery and exceptions, which stay on the main CPU?
- How do random small-block IOPS hold up at 512 bytes and at 4 KB?
- Is firmware tooling mature enough to ship reliable updates for a 100-plus-core controller?
Bottom line
SK hynix's pitch is that SSD controllers should be more like a well-run kitchen than one overworked chef. A pile of cheap, single-purpose cooks handles repetitive prep, and the head chef deals with the emergencies. The logic is sound, and Phison's micro-core approach shows it has precedent. But the per-core figures are the company's own, and there is no named drive, benchmark or ship date. Treat the PCIe 8.0 framing as a roadmap rationale, not a product spec.
References
- Samsung's archrival wants to spread 100 special CPUs in every SSD controller ahead of PCIe 8.0, High Bandwidth Flash onslaught TechRadar · 2026-10-24T21:20:00+00:00
- SK hynix H-TPU Architecture Moves SSD Controller Work Onto Tiny Specialised Cores guru3d.com
- SK Hynix Unveils H-TPU Controller Architecture with Over 100 Small Processors < Semiconductor < 기사본문 - The Elec Inc. thelec.net