SK hynix has shipped samples of its 12-stack HBM4E high-bandwidth memory to major AI customers, marking an early step toward the company’s next generation of HBM supply for AI training and inference platforms.

The 12-layer HBM4E is rated for a maximum data rate of 16 Gbps per pin. SK hynix also states that power efficiency is up more than 20 percent versus previous models. The company is positioning those gains as headroom for both AI training and inference workloads, where memory bandwidth and power can directly constrain accelerator utilization and rack-level power budgets.

SK hynix says HBM4E reduces data-transfer latency through a “latest interface” and design optimizations, while maintaining stable operation in high-bandwidth environments. For data center operators, the practical point is simple: as accelerators scale, memory can become a limiter long before compute does, and incremental bandwidth-per-watt improvements can matter just as much as peak bandwidth when you’re trying to stay inside a fixed kW/rack envelope.

On the packaging side, SK hynix is using its Advanced MR-MUF technology in the HBM4E stack. The company specifies 48 GB capacity in a 12-layer stack, and it says heat resistance is improved by 17 percent compared to HBM4. That thermal detail is worth watching: HBM is tightly coupled to accelerators, and a little more margin at the memory stack can ease system-level cooling and reliability constraints when everything is pressed up against temperature limits.

SK hynix said it delivered the samples on schedule and plans to work with partners toward mass production. “We will work closely with partners for mass production in a timely manner,” the company said. Ahn Hyun, President and Chief Development Officer, added, “Through close collaboration with our partners, we will deliver the value needed in the market while reinforcing our technology leadership as a full-stack AI memory creator.”

Source: SK hynix