🇰🇷 한국어 🇺🇸 English 🇯🇵 日本語 🇪🇸 Español 🇨🇳 简体中文
Published: 2026.11.22 (Sun)
Economy

Memory-centric design becomes vital as AI enters the physical world

As artificial intelligence transitions from data centers to physical environments like robots and autonomous vehicles, the focus of competition is shifting from pure computing power to memory-centric design.

Memory-centric design becomes vital as AI enters the physical world
A graphic image featuring a semiconductor chip and text themed around the AI ecosystem. (Photo=SK hynix)

As artificial intelligence moves into robots and autonomous vehicles, the focus is shifting from computing power to memory-centric design to manage power and latency constraints. To address these physical limits, SK hynix aims to provide full-stack solutions including HBM, PIM, and HBF to optimize data movement for On-device AI.

Artificial intelligence (AI) is entering the era of "Physical AI," where it moves beyond massive data centers to interact with the real world through robots, autonomous vehicles, and smart glasses. As AI must make real-time decisions and movements in physical environments, the core challenge is shifting from traditional competition centered on computing power to "memory-centric design," which focuses on how data is stored and efficiently moved.

Professor Nam Sung Kim of the University of Illinois Urbana-Champaign (UIUC) analyzed that as AI expands into the real world, physical constraints such as power consumption, heat generation, latency, and product size become the decisive variables determining performance. Unlike data centers, edge or mobile environments have limited battery capacity and space, making structural innovation essential to reduce the distance between computing devices and memory while minimizing data movement.

Physical limits faced by AI outside data centers and the evolution of algorithms

Until now, high-performance AI has been optimized for data center environments equipped with large-scale GPU clusters and massive power supplies. However, the situation is different for On-device AI environments, such as smartphones and wearable devices. As devices become smaller, the available energy decreases and the space to handle generated heat narrows. In particular, tasks sensitive to latency, such as robot movements or driving decisions for vehicles, require "low latency" characteristics that allow data to be processed immediately near the site instead of being sent to external servers.

To overcome these constraints, AI algorithms are also changing. Representative examples include "quantization" techniques that reduce model size to lower memory burdens, and "TurboQuant," which reduces memory usage during the inference process. Additionally, architectures have emerged that reduce the computing burden by utilizing memory lookups instead of repetitive calculations, such as the "Engram" method, which pre-stores frequently used information in memory to be retrieved when needed. This demonstrates that the evolution of algorithms ultimately depends on memory capacity and data supply capabilities.

Integrated design of memory hierarchies from HBM to HBF

To realize the performance of Physical AI, a "memory-centric architecture" must be built to optimize the location and movement paths of data. To achieve this, various memory technologies are deployed hierarchically according to their respective roles. First, custom High Bandwidth Memory (HBM) and 3D stacked DRAM serve to widen data paths and reduce latency by placing memory as close to the computing devices as possible. By adding Processing-in-Memory (PIM) technology, which performs some computations within the memory, energy consumption caused by data movement can be drastically reduced.

The role of the storage layer, responsible for storing and supplying data, is also increasing. Sensor data from robots or autonomous driving maps are stored in high-capacity SSDs. Recently, HBF (High Bandwidth Flash), a next-generation memory technology that combines the high capacity of NAND with high transfer speeds, has been discussed to bridge the gap between HBM and SSDs. Furthermore, CXL (Compute Express Link) technology provides the foundation for flexibly expanding memory resources by high-speed connecting processors with various memory devices.

System-level "full-stack solutions" will determine AI performance

Ultimately, the deciding factor in the era of Physical AI lies beyond the performance of individual memory products; it depends on how the entire memory hierarchy is combined to match specific workloads and device characteristics. This is because the memory specifications required for large-scale AI for training are completely different from those for On-device AI, which prioritizes low power and low latency.

In response to these changes, SK hynix aims to provide an integrated "full-stack AI memory solution" from a system perspective, encompassing HBM, PIM, CXL, SSD, and HBF. The strategy is to strengthen its role as an "Architecture Partner" that designs memory combinations optimized for a customer's architecture and workload, moving beyond being a mere manufacturer that supplies memory components. As AI expands into the real world, memory is expected to function as a core design axis that determines the response speed and efficiency of a system, rather than just a simple storage device.

Source: news.skhynix.co.kr

#Physical AI #SK hynix #HBM #PIM #CXL #memory architecture #Nam Sung Kim #UIUC
L
Lim Sangwoo
TrendBiz · Reporter

Covers Economy for TrendBiz, and also writes about Company News and Finance.

More by this reporter ›
Copyright ⓒ TrendBiz All rights reserved. Unauthorized reproduction, redistribution, and AI training prohibited.

0Comments

Comments are currently disabled.

Be the first to comment.