Keynote at 'The Six Five Summit' in the U.S.

"Samsung sells bandwidth, not bits"

"'Memory-centric computing' is becoming crucial"

Samsung Electronics has presented memory bandwidth as a key competitive edge in the era of artificial intelligence (AI), and predicted that the advent of next-generation high-bandwidth memory (HBM), known as zHBM, will occur earlier than previously expected, by 2029.


Speaking as a keynote speaker at 'The Six Five Summit 2026', a U.S. information technology (IT) conference held online on the 27th (local time), Sangyeon Jo, Head of the Americas Division of Samsung Electronics Semiconductor (Device Solutions) Division (DSA), shared these insights.


During a 30-minute talk titled "Redefining AI Limits: Samsung's Memory, Power, and System Design," Jo pointed out, "The industry still does not fully understand that the 'tokens per second required by an organization'—an extremely important factor for AI inference services—is determined by memory bandwidth."


He explained that memory bandwidth, which shows how quickly data can move between memory and computing units, determines the number of users, the number of requests, response speed, context retention, and throughput. The level of service an AI system can provide, as well as the cost required to implement it, both depend on memory bandwidth.


Sangyeon Cho, Head of the Americas Corporation (DSA) for Samsung Electronics Semiconductor (DS) Division. Screenshot of The Six Five Summit X account.

Sangyeon Cho, Head of the Americas Corporation (DSA) for Samsung Electronics Semiconductor (DS) Division. Screenshot of The Six Five Summit X account.

View original image

Jo added, "Samsung does not sell 'bits' (the units of data); rather, we sell 'bandwidth.' Four years from now, the most successful applications will not be those with the largest memory capacity, but those that figure out how to effectively utilize their memory bandwidth."


He further suggested 'memory-centric computing', proposing to set the required bandwidth starting from memory and then seek efficient ways to leverage it. Traditionally, computing resources have been decided first, followed by memory capacity, packaging, and power issues, but going forward, he stressed that this sequence should be reversed with memory considered at the forefront.


As an example of a product aimed at increasing memory bandwidth, Jo highlighted Samsung's recently introduced 'zHBM.' Distinct from conventional approaches where high-bandwidth memory (HBM) is connected alongside the AI accelerator (xPU), zHBM adopts an architecture where the HBM is stacked directly atop the accelerator, minimizing data transfer distance and maximizing bandwidth.


He stated, "Samsung is currently defining the specifications for the first generation of zHBM, and based on this, related future products are expected to launch after 2029." This is one year ahead of earlier industry forecasts, which expected commercialization around 2030.


However, he also noted that resolving the heat dissipation challenge remains a key technical issue for this product's development. To address this, technologies such as hybrid copper bonding (HCB) are being applied.


Alongside memory bandwidth, Jo pointed to advanced packaging as another area of growing importance. As process miniaturization alone reaches its limit in meeting the performance and power efficiency demands of AI, the role of packaging technology—integrating multiple dies, memory, and diverse process nodes into a single system—is becoming ever more vital.


"Some companies outsource their base die and manage the handoff between process steps," he emphasized. "In contrast, Samsung directly manages the integration at every stage—system-level, chip-level, and memory-level—to realize design objectives." For example, in the case of HBM4, Samsung designs the base die in-house using an advanced 4-nanometer process, and manages assembly and packaging in a one-stop approach, optimizing power efficiency, signal integrity, performance, and bandwidth across the entire system.



Meanwhile, Jo also said that Samsung plans to operationalize its second semiconductor fabrication plant, Taylor Fab 2, in Taylor, Texas, within the next few years, thereby expanding its production capacity for advanced process nodes including 2 nm and supporting its U.S. customers.


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing