SK hynix unveiled its New Spectrum roadmap for AI infrastructure components at the AI Infra Summit 2026 held in Santa Clara from September 15th to 17th.

SK hynix participated in the AI Infra Summit 2026, held September 15–17 in Santa Clara, California. The three-day conference examined emerging technology trends and explored future directions across five tracks: Data & Models, Compute, Data Movement, AI Data Center, and Physical AI. With major global technology firms showcasing their latest innovations, the exhibition floor remained highly active throughout the event.

Sustaining AI’s rapid growth requires continuous innovation in core infrastructure such as servers and data centers. Reflecting this priority, the summit attracted approximately 6,000 attendees—double the turnout from the previous year. In response, SK hynix doubled its booth footprint and showcased an extensive portfolio of next-generation memory solutions, reinforcing its leadership position in AI infrastructure.

Operating under the “New Spectrum” theme, the SK hynix booth highlighted its latest AI memory technologies and strengthened its brand presence. The layout was designed for immediate comprehension: the rear wall featured a graphical representation of the company’s technology vision, the left side showcased PIM and SALT-KV technologies, and a detailed HBF structural model stood prominently at the entrance.

The HBF display featured an introductory video alongside a physical structural model. As the AI industry shifts from training to inference, demand is growing for a new memory tier positioned between HBM, which delivers high bandwidth, and SSDs, which offer massive storage capacity. HBF (High Bandwidth Flash) is a next-generation NAND solution that leverages TSV technology, similar to HBM, to combine the high capacity of traditional storage with high-speed data transfer. Recognized as a foundational technology for the inference-driven era of Agentic AI, the HBF exhibit drew the highest visitor engagement. Attendees frequently watched the full introductory video and spent significant time studying the product mock-up to understand its architecture.

The PIM (Processing-in-Memory) exhibit displayed SK hynix’s AiM chip, the AiMX (AiM-based Accelerator) card, and a fully integrated server, alongside a hands-on demonstration of an enhanced LLM service. PIM embeds computational functions directly into memory to minimize data movement between the processor and memory, addressing the “Memory Wall” bottleneck while boosting both performance and power efficiency. Visitors showed strong interest in the technology’s potential to streamline AI workloads, expressing clear anticipation for its near-term deployment.

The SALT-KV (Semantic-Aware Lifecycle Tiering for KV Cache) display featured a server equipped with SK hynix’s eSSD and an adjacent monitor that visualized real-time data movement within an AI system. SALT-KV partitions an LLM’s KV cache into context-based segments, evaluating each for reuse value and storage cost before dynamically assigning it to the optimal tier among HBM, DRAM, and SSD. This approach maximizes utilization within constrained memory resources. The exhibit also hosted an interactive demo where attendees used an LLM powered by SALT-KV to build and run a simple game, consistently drawing crowds eager to test the solution live.

Complementing the exhibition, SK hynix reinforced its commitment to meeting evolving customer demands through a dedicated session presentation. On the second day of the conference, Vice President and Head of Solution AT Lim Eui-cheol addressed the Data Movement track under the theme “Beyond One-Size-Fits-All: PIM, HBF and More for the New Spectrum of AI Serving,” outlining ongoing R&D advancements in next-generation memory technologies.

“As AI workloads grow more diverse—ranging from ultra-low-latency agent services to long-context applications where cost efficiency is critical—the conventional GPU-HBM architecture alone is increasingly insufficient,” Lim stated. “To address these shifting demands, SK hynix is developing integrated hardware and software solutions, including PIM, HBF, and SALT-KV.” He further noted, “HBF combines high bandwidth and large capacity to optimize long-context workloads, while PIM enhances performance and efficiency for the fast-decoding demands of premium services. Both represent strong pathways to surpass current architectural limits. SALT-KV, which boosts processing efficiency by accounting for KV cache characteristics and reusability, also holds significant potential to advance AI system evolution.”

Through its participation in the AI Infra Summit 2026, SK hynix showcased next-generation AI memory solutions—including PIM, HBF, and SALT-KV—and articulated its strategy to lead innovation in AI infrastructure. The company remains committed to advancing its technological capabilities to meet the rapidly evolving demands of the AI era.

**Reference Definitions**

* HBM (High Bandwidth Memory): A high-value, high-performance product that vertically connects multiple DRAM chips to increase capacity and dramatically improve data processing speed. HBM has been developed in the following order: first generation (HBM), second generation (HBM2), third generation (HBM2E), fourth generation (HBM3), fifth generation (HBM3E), and sixth generation (HBM4).

* TSV (Through Silicon Via): A technology that connects chips vertically through electrodes that pass vertically through the silicon substrate.

* AiM (Accelerator-in-Memory): A next-generation solution that embeds a processor’s computational functions into memory.

* LLM (Large Language Model): An artificial intelligence model that learns from big data to generate and understand natural, human-like sentences. By dramatically increasing the scale of its parameters, it acquires advanced language-processing capabilities such as complex context comprehension, reasoning, and translation.

* KV Cache (Key Value Cache): Temporary memory in which a large language model-based AI agent stores previously computed results when generating text. The longer the conversation with a user and the more tasks there are to process, the more exponentially it grows—so the memory capability to store and process it efficiently has a major impact on an AI agent’s performance.