Samsung, SK hynix to unveil 'thinking memory' tech at Hot Chips as AI bottleneck shifts focus

It is, quite literally, a hot summer. California heats up even further around this time every year — and this week, two major semiconductor conferences are to blame: FMS (Future of Memory and Storage) and the well-known Hot Chips, both held in the state. With interest in semiconductors running higher than ever this year, the temperature feels a few degrees warmer still.

FMS is the world's largest independent conference dedicated to memory and storage technology, bringing together chipmakers such as Samsung Electronics, SK hynix and Micron alongside storage and data center companies to unveil next-generation memory and storage technologies and roadmaps — covering HBM (high-bandwidth memory), DRAM, NAND, SSDs (solid-state drives) and CXL (Compute Express Link). This week, both Samsung Electronics and SK hynix revealed their next-generation memory technologies, offering a clearer picture of where each company is headed beyond HBM. More on that shortly.

The conference to cover first this Friday is Hot Chips. Launched in 1989 and held every August in Silicon Valley, it is the premier conference for high-performance semiconductor and computing architecture. Nvidia, AMD, Intel, IBM and Arm are among the major companies that use the event to present next-generation CPUs (central processing units), GPUs (graphics processing units), AI accelerators, and memory and packaging technologies. What sets Hot Chips apart is its focus on real products and implementable technologies rather than academic papers.

In the industry, CES — the world's largest IT show held every January in the United States — serves as a gauge of the year's trends. Hot Chips plays a similar role as a weathervane: its program lineup offers a fairly clear read on what the industry considers most important.

Looking at Hot Chips themes over the past three years: in 2024, Nvidia presented on its Blackwell accelerator and AMD on its MI300X, with hardware design using AI as the central thread. In 2025, the focus shifted from individual chips to AI data centers, with Nvidia, AMD and Google discussing rack architecture, GPUs, TPUs (tensor processing units) and networking. Those concepts have since become reality.

This year's theme is striking. The very first tutorial is titled "Memory Technology," and a dedicated memory session has been formally added to the program. It is yet another signal that memory is the true bottleneck in AI.

What makes this year even more notable is that Samsung Electronics and SK hynix will take the stage together for the first time in a decade. The last time both companies presented at Hot Chips was in 2016, when each spoke about HBM.

In 2016, Samsung Electronics presented under the title "The Future of Graphic and Mobile Memory for New Applications," laying out a vision for HBM and graphics memory. SK hynix, meanwhile, presented "The Era of High Bandwidth Memory," unveiling the direction for next-generation HBM.

Looking back at those presentations a decade later is a remarkable experience. Samsung Electronics introduced the concept of PIM — processing-in-memory, a technology that embeds computation inside the memory chip itself — alongside a broader vision for expanding HBM's use cases. SK hynix outlined the path from HBM2 toward HBM3, with improvements in bandwidth, capacity and power efficiency. Nearly everything sketched out on those blueprints a decade ago has since become reality.

This year promises to be even more compelling, offering a glimpse of what both companies envision for memory over the next decade. In the opening "Memory Technology" session, Micron researcher Raghu Sriiramaneni will present first on "Evolving memory architectures for AI."

Following that, Samsung Electronics' Han Sang-wook will present on "HBM Base Die: How will HBM evolve in the future by utilizing the logic process?" and SK hynix America Vice President Lee Jae-sik will speak on "Advanced Packaging for High Bandwidth Memory."

Samsung Electronics is putting the HBM base die front and center at this year's Hot Chips. In-house base die design and production capabilities are widely cited as one of the key factors behind the company's ability to reclaim market leadership in HBM4. (The base die sits at the very bottom of an HBM stack and can be thought of as the chip's control tower.) The strategy appears aimed at translating Samsung's strengths as an integrated device manufacturer — with capabilities spanning memory, logic design, foundry and advanced packaging — into a competitive edge in next-generation HBM.

Particularly notable is the explicit mention of "logic process" in Han's presentation title. It signals that HBM's future competitiveness will no longer be determined by DRAM process technology alone, but will increasingly depend on the logic process and design capabilities applied to the base die.

Samsung Electronics' emphasis on the base die also suggests the company will push harder into custom HBM. Where memory products were once largely uniform regardless of whether they went into Nvidia or AMD systems, the growing diversity of accelerator architectures — GPUs, TPUs and beyond — has fueled demand for tailored HBM. The idea is to supply memory optimized for each customer: HBM built for Nvidia, HBM built for Google, HBM built for AMD. The approach carries the added benefit of stronger customer lock-in.

Samsung Electronics' other presentations fill out its memory roadmap. On Aug. 25, the company will present on LPDDR5X (low-power DRAM) combined with PIM (processing-in-memory). PIM is sometimes described as memory that mimics the brain: by embedding computation directly inside the memory chip, it can handle simple operations on the spot rather than shuttling data to a processor.

The session title is "Samsung LPDDR5X-PIM: World's First LPDDR based Processing in Memory Solution for AI Inference."

Until now, PIM has been concentrated in high-performance memory for data centers, such as HBM-PIM. Extending it to LPDDR5X — the low-power memory found in the smartphones and AI PCs people use every day — marks a significant expansion. Where conventional chips require a CPU or GPU to pull data from memory, process it and write results back, PIM handles some of those operations inside the memory itself, passing only the necessary output to the CPU or NPU. The result is less data movement, higher performance and better power efficiency.

A joint presentation by Samsung Electronics and XCENA — a South Korean fabless semiconductor startup — on CXL points in the same direction. CXL is a next-generation interconnect technology that enables CPUs, GPUs and memory to communicate seamlessly. Connecting a CXL-enabled CMD-D (CXL memory module) allows memory capacity to be expanded substantially.

Samsung Electronics' direction is clear: rather than moving data all the way to a CPU or GPU for processing, the company wants to place computation closer to the memory itself, handling some tasks on the spot. The goal is to evolve memory's role from a device that simply stores data into one that also assists with computation.

SK hynix's emphasis on packaging is equally worth noting. The company combines TSMC's advanced logic process with CoWoS (chip on wafer on substrate), TSMC's advanced packaging technology — a combination that leaves SK hynix with considerable dependence on TSMC. As HBM advances through successive generations, challenges around heat dissipation, warping, thickness and yield grow more acute, making back-end packaging technology increasingly critical. SK hynix is also exploring hybrid bonding as it prepares for an era of HBM stacks with 16 or more layers.

The key takeaway from this year's Hot Chips is the elevated status of memory and its evolution beyond simple storage and supply — toward a role that shares in the work of computation. There has long been a tendency to dismiss Korea as merely a "chip factory." What both companies are signaling now is an ambition to move from being simple chip suppliers to becoming core design partners in AI systems. Customers will no longer be able to treat memory as a commodity, and for the chipmakers themselves, this looks like a pivotal moment to capture meaningfully higher margins.

For the new technologies from Samsung Electronics and SK hynix introduced at FMS (Future of Memory and Storage) — mentioned at the top — see the related articles below. Stay cool out there.