Search AI Chips News Today

Chipmakers

SK hynix HBM4: 2,048-bit interface, 12-layer stacks, 40% power gain

SK hynix HBM4 doubles I/O to 2,048 bits, hits 10 Gbps, and cuts power 40% versus HBM3E. We break down the architectural impact for AI accelerators.

Precision Wafer Processing in a Cleanroom. Photo: ResearchFrontier (CC BY 4.0, Wikimedia Commons).

SK hynix has completed development of HBM4 and declared it ready for mass production, shifting its target to the second half of 2025. The 12-layer stacks, which first went to key customers including Nvidia six months earlier, are the first HBM to process over 2 TB of data per second. The move accelerates a schedule that originally pointed to early 2026, after Nvidia CEO Jensen Huang requested 12-layer HBM4 six months ahead of plan. For hardware engineers and infrastructure buyers, the headline is not just speed: it is a wider per-stack interface, a logic base die built on TSMC's advanced logic process, and a claimed 40% power-efficiency gain over HBM3E that changes how accelerator bandwidth budgets and data centre power envelopes are calculated.

The 2,048-bit interface and what it does to bandwidth budgets

HBM4 doubles the number of I/O terminals to 2,048, up from 1,024 on HBM3E. JEDEC finalized the HBM4 specification on April 16, setting a 2,048-bit-wide interface and an 8 Gb/s maximum data rate that yields 2.048 TB/s per stack. SK hynix says its parts deliver 10 Gbps, exceeding the JEDEC operating-speed standard, and claims a 60% speed gain from doubling channels from 1,024 to 2,048. The company also claims a potential 69% AI performance boost.

For accelerator designers, the practical effect is that a single HBM4 stack now supplies roughly the bandwidth that previously required multiple HBM3 stacks. Semiconductor Engineering notes that four HBM3 stacks at 6.4 Gb/s deliver 3.2 TB/s, while the bandwidth ceiling to a single GPU or accelerator is around 10 TB/s. HBM4 moves per-stack bandwidth to 2.048 TB/s at JEDEC rates, so fewer stacks can saturate the die-to-memory interface. That changes the trade-off between stack count, interposer area and power. It also raises the bar for the memory controller: Rambus lists a 10 Gb/s maximum data rate for its HBM4 controller, and the interface stays backwards compatible with existing HBM3 controllers, which lowers the integration risk for teams already shipping HBM3E designs.

12-layer stacking, 16-high ceilings and the base-die logic shift

SK hynix developed 12-layer HBM4 as a world-first in September 2025. JEDEC's specification supports stacks up to 16-high with 32 Gb die densities, yielding 64 GB in a single stack. Industry reports link SK hynix 12-layer HBM4 to Nvidia's Vera Rubin GPU, and separate reports say the Rubin Ultra dual GPU will stack 16 HBM4 chips. Those reports are unconfirmed, but they indicate the direction of travel: more layers per stack and more stacks per package.

The base die is the other architectural change. HBM4 is the first product of SK hynix's manufacturing collaboration with TSMC, which supplied an advanced logic process for the HBM4 base die. Moving the base die to a logic process puts more of the interface and control logic inside the stack rather than on the host die. SK hynix pairs this with its MR-MUF packaging, which injects and hardens liquid protective material between stacked chips, and fifth-generation 1bnm DRAM. JEDEC also adds Directed Refresh Management for RAS and row-hammer mitigation, a detail that matters as stacks grow taller and densities rise. HBM3E devices connect via 1,024 data wires, about 1,700 with command and address, and the denser HBM4 interface still requires a silicon interposer to route the wiring. HBM4 implementation and manufacturing costs exceed GDDR7's 2D-memory architecture, so the economics favour accelerators that can amortise the interposer and packaging across high-volume training and inference fleets.

Power efficiency and the data centre envelope

SK hynix claims over 40% better power efficiency than HBM3E. JEDEC lists a 0.7V lowest VDDQ option and a 1.05V highest VDDC option, giving board designers more room to trade voltage against frequency. In a cluster of 100,000 GPUs, memory power is no longer a rounding error; a 40% efficiency gain on the memory subsystem can free headroom for compute or reduce cooling load. The 12-layer HBM4 is also the first HBM to process over 2 TB of data per second, which SK hynix equates to about 400 full-HD movies in one second.

The context is a market SK hynix leads. TrendForce put its 2024 global HBM share at 52.5%, against 42.4% for Samsung and 5.1% for Micron. An analyst projects SK hynix's HBM share in the low 60% range in 2026. Micron and Samsung plan HBM4 launches in 2026, behind SK hynix. SK hynix was first to supply fifth-generation HBM3E to Nvidia in 2024, and HBM3E with 288 GB will pair with Blackwell Ultra later this year. At SC25 in St. Louis from November 16 to 21, SK hynix displayed 12-layer HBM3E with Nvidia's GB300 Grace Blackwell GPU, alongside DDR5 RDIMM and MRDIMM modules on the 1c node, a 245 TB PS1101 E3.L SSD on 321-layer QLC NAND, and a CXL pooled memory machine for distributed LLM inference.

Practical takeaways for architects and buyers

  • Recount your stack budget. At 2.048 TB/s per stack under JEDEC rates, fewer HBM4 stacks can reach the roughly 10 TB/s ceiling to a single accelerator, freeing interposer area and package power.
  • Plan for 16-high and 64 GB stacks. JEDEC supports 16-high stacks with 32 Gb dies, so capacity per stack can double versus today's 12-layer parts; confirm vendor qualification before committing.
  • Treat the logic base die as a design variable. TSMC's advanced logic process for the HBM4 base die means the memory stack now carries more interface logic; validate controller compatibility, since HBM4 stays backwards compatible with HBM3 controllers.
  • Budget the 40% power gain into the rack. Use the efficiency claim to offset higher stack counts, but model VDDQ and VDDC options (0.7V to 1.05V) against your cooling envelope.
  • Watch the unannounced details. SK hynix has not disclosed volume ramp specifics, customer qualification status beyond readiness for mass production, or confirmation that 12-layer HBM4 ships with Vera Rubin. Micron and Samsung timing beyond 2026 is also open.

For deeper context on how memory bandwidth shapes accelerator design, see our AI chipmakers coverage hub and why memory matters for AI chips. JEDEC's published HBM4 standard remains the reference point for interface and timing details at the JEDEC Solid State Technology Association, and EE Times reported the SK hynix HBM4 production milestone.