SK hynix argues memory is becoming a core design variable in physical AI

November 17, 2026

SK hynix argues memory is becoming a core design variable in physical AI

Image: ChatGPT

SK hynix says the next phase of AIartificial intelligenceAI stands for artificial intelligence will be defined less by compute alone than by how efficiently systems store, move, and process data. As workloads move toward inference and agentic AIartificial intelligenceAI stands for artificial intelligence, memory subsystem architecture becomes a central part of performance, especially when AIartificial intelligenceAI stands for artificial intelligence is deployed outside datacenters and into physical devices that must react in real time.

The company frames physical AIartificial intelligenceAI stands for artificial intelligence as a design problem in which latency, power consumption, heat dissipation, and package size all matter at once. Robots, vehicles, industrial drones, smart glasses, and wearables cannot rely on datacenter-style resources such as large GPUgraphics processing unitA GPU is a parallel graphics processor widely used for AI acceleration. clusters, abundant board area, and heavy cooling. In these environments, processing data close to where it is generated becomes essential.

That shift also changes algorithm design. Model compression reduces model size, while quantization lowers the number of bits used to represent parameters and cuts storage and transfer demands. SK hynix also points to TurboQuant as a quantization approach aimed at reducing memory requirements during inference without sacrificing response quality.

Another example is Engram, which moves away from repeated layer-by-layer computation and instead retrieves frequently used patterns from a conditional memory lookup table. This reduces recomputation, but it shifts the bottleneck toward memory capacity and data delivery, because more information must be stored and fetched quickly.

On the hardware side, the article groups several memory technologies by their role in the hierarchy. HBM1. human body modelESD test model that simulates a person discharging static electricity into a component.2. High Bandwidth MemoryHBM is a stacked DRAM architecture designed to provide very high memory bandwidth. uses vertically stacked DRAM1. dynamic random access memoryMemory technology or system for temporarily or permanently storing digital information.2. dynamic random-access memoryDRAM is a volatile memory technology dies and a wide interface to deliver high bandwidth in a small footprint. 3D-stacked DRAM1. dynamic random access memoryMemory technology or system for temporarily or permanently storing digital information.2. dynamic random-access memoryDRAM is a volatile memory technology places DRAM1. dynamic random access memoryMemory technology or system for temporarily or permanently storing digital information.2. dynamic random-access memoryDRAM is a volatile memory technology directly on top of logic chips or SoCs to shorten the distance between memory and compute. Custom DRAM1. dynamic random access memoryMemory technology or system for temporarily or permanently storing digital information.2. dynamic random-access memoryDRAM is a volatile memory technology adapts capacity, bandwidth, and power characteristics to specific workloads and form factors.

Processing-in-memory moves selected operations closer to the data itself to reduce unnecessary transfers, while CXLCompute Express LinkA high-speed interface standard used here to pool memory across servers and GPUs. provides a high-speed serial interface for connecting processors, accelerators, and external memory resources. For data volumes that do not fit in DRAM1. dynamic random access memoryMemory technology or system for temporarily or permanently storing digital information.2. dynamic random-access memoryDRAM is a volatile memory technology or HBM1. human body modelESD test model that simulates a person discharging static electricity into a component.2. High Bandwidth MemoryHBM is a stacked DRAM architecture designed to provide very high memory bandwidth., SSD1. Solid State Drivenon-volatile electronic storage medium, usually a hard disk drive in computers2. Solid-State DriveA non-volatile storage device used in the rack demo to hold full KV caches.s based on 3D NAND1. Not ANDA logic gate that outputs 0 when all inputs are 1, and outputs 1 in all other cases2. negative-ANDNAND denotes a logic operation and is also the name of a widely used flash memory type. store sensor logs, map data, and long-term operating data. HBF1. High Bandwidth FlashA NAND-based concept that combines storage capacity with higher bandwidth.2. High-Bandwidth FlashA memory concept under development that is intended to relieve bandwidth and capacity bottlenecks in AI inference systems. is presented as a new layer intended to combine NAND1. Not ANDA logic gate that outputs 0 when all inputs are 1, and outputs 1 in all other cases2. negative-ANDNAND denotes a logic operation and is also the name of a widely used flash memory type. capacity with higher transfer speeds for frequently accessed inference data such as model weights.

The broader message is that AIartificial intelligenceAI stands for artificial intelligence memory is moving from a product-by-product comparison to a system-level design problem. SK hynix presents its full-stack memory strategy as a combination of HBM1. human body modelESD test model that simulates a person discharging static electricity into a component.2. High Bandwidth MemoryHBM is a stacked DRAM architecture designed to provide very high memory bandwidth., DRAM1. dynamic random access memoryMemory technology or system for temporarily or permanently storing digital information.2. dynamic random-access memoryDRAM is a volatile memory technology, SSD1. Solid State Drivenon-volatile electronic storage medium, usually a hard disk drive in computers2. Solid-State DriveA non-volatile storage device used in the rack demo to hold full KV caches.s, HBF1. High Bandwidth FlashA NAND-based concept that combines storage capacity with higher bandwidth.2. High-Bandwidth FlashA memory concept under development that is intended to relieve bandwidth and capacity bottlenecks in AI inference systems., CXLCompute Express LinkA high-speed interface standard used here to pool memory across servers and GPUs., and PIM1. Processing-In-MemoryAn approach that performs some computation close to or inside memory.2. processor-in-memoryPIM describes computing directly inside memory to reduce data movement., with Memory Library capabilities intended to help tailor configurations to customer architectures and workloads. The company positions itself not only as a component supplier, but as an architecture partner for AIartificial intelligenceAI stands for artificial intelligence systems whose performance depends on the entire data lifecycle.

As AIartificial intelligenceAI stands for artificial intelligence spreads beyond datacenters, memory is becoming a decisive part of how quickly systems respond and how efficiently they operate. The article argues that the future evolution of AIartificial intelligenceAI stands for artificial intelligence will depend increasingly on how well memory architecture is matched to the physical constraints of the target device.

Source: SK hynix

More background: How does it work?

← All posts