SK hynix unveils 'beyond HBM' technologies at US AI infrastructure summit

SK hynix unveiled a wave of next-generation technologies set to go beyond high-bandwidth memory at the AI Infra Summit 2026, citing growing limitations of the conventional GPU-HBM architecture as AI capabilities continue to advance.

The chipmaker participated in the AI Infra Summit 2026, held Tuesday through Thursday in Santa Clara, California, presenting future AI memory technologies under the theme "New Spectrum." This year's event drew about 6,000 attendees — roughly double last year's turnout — and SK hynix doubled the size of its booth from the previous year.

To address the expanding AI inference market, SK hynix pushed next-generation memory solutions including high-bandwidth flash (HBF), processing-in-memory (PIM) and SALT-KV, signaling a strategy to broaden its business beyond HBM into memory and storage solutions optimized for specific AI workloads. As the AI market rapidly shifts from large-scale training toward real-time inference and AI agents, competition in the memory industry is widening beyond simple bandwidth gains to encompass capacity, power efficiency and overall data-processing approaches.

SK hynix placed particular emphasis on HBF, a solution built by stacking layers of NAND flash. As AI's paradigm shifts from training to inference, demand is growing for a new memory tier sitting between the high bandwidth of HBM and the large storage capacity of solid-state drives.

HBF applies through-silicon via (TSV) technology — the same approach used in HBM — to deliver both the large capacity of conventional SSDs and high bandwidth at the same time. It is drawing attention as a key next-generation NAND solution poised to define the era of inference-driven agentic AI.

SK hynix also demonstrated its PIM technology, displaying the "AiM" chip — a memory device with built-in processing capability — alongside the "AiMX" accelerator card, which houses multiple AiM chips, and a server equipped with AiMX cards.

PIM embeds computing functions inside the memory itself to reduce data movement between the processor and memory, easing the bottleneck known as the memory wall while improving both computational performance and power efficiency.

The company also introduced SALT-KV, a technology designed to tackle the surging KV cache storage demands that arise during AI inference. It segments KV cache by context unit, evaluates each segment's reuse value and storage cost, and distributes the data across HBM, DRAM and SSDs accordingly.

The goal is to use limited high-performance memory more efficiently, cutting the overall cost and improving the performance of AI systems. SK hynix demonstrated SALT-KV in operation through a server equipped with an enterprise SSD.

Im Eui-cheol, head of SK hynix's Solution AT division, shared the company's research and development progress in next-generation technologies in a presentation titled "Beyond One-Size-Fits-All: PIM, HBF and More for the New Spectrum of AI Serving."

"As AI workloads diversify, the existing GPU-HBM architecture alone is becoming insufficient to meet new requirements," Im said. "We are preparing PIM, HBF and SALT-KV as new solutions spanning both hardware and software to match the evolving workload landscape."

He added that HBF, combining high bandwidth and large capacity, can boost the efficiency of AI services that must handle long contexts, while PIM — which improves performance and efficiency for fast-decoding workloads in premium services — offers a strong alternative that can overcome existing limitations. "SALT-KV, which enhances processing efficiency by accounting for the characteristics and reusability of KV cache, also holds strong potential to drive the evolution of AI systems," he said.

SK hynix said it would "continue strengthening its technological competitiveness to stay one step ahead in responding to the diverse demands of the AI era."