Samsung Patents a Storage Device That Trains AI Models Internally

By Patentlyze Team · Updated May 22, 2026

Instead of shipping raw data from storage to a CPU or GPU for AI model training, Samsung's patent proposes letting the storage device do the math itself — a classic 'bring compute to the data' move that could quietly reshape how machine learning hardware is architected.

What Samsung's in-storage AI training actually does

Imagine every time you want to train an AI model, your computer has to haul enormous amounts of data from your hard drive all the way to the processor — over and over, billions of times. That data movement is slow, energy-hungry, and increasingly the thing that slows modern AI training down.

Samsung's patent flips that assumption. Instead of moving data to the compute, it proposes doing the AI training calculations inside the storage device itself. The drive receives a signal from the host system saying "update your model parameters," crunches the numbers using the data already sitting right there, and sends back just the results — not the raw data.

For you as an end user, this could eventually mean faster, cheaper AI training — because less data has to travel across the system bus, and the host processor gets to focus on higher-level tasks rather than shuffling numbers around.

How the storage device calculates DNN parameters on-device

This patent describes a computational storage architecture where a storage device — think an SSD or similar flash-based hardware — takes on active responsibility for calculating deep neural network (DNN) model parameters during training, rather than passively serving raw data to an external processor.

The workflow has three steps:

• The host sends input data to the storage device (or the device already holds it).
• When the host issues a parameter updating request — essentially saying "it's time to update the model" — the storage device computes the DNN parameters locally using that data. The patent abstract specifically calls out momentum and variance as values being calculated, which are key components of optimizers like Adam (a popular algorithm used to update neural network weights during training).
• The storage device transmits at least a portion of the calculated parameters back to the host — meaning it can selectively return only what the host needs, not the entire dataset.

This is a form of near-data processing (NDP) or computational storage — a long-studied idea in computer architecture where logic is embedded close to where data lives to avoid the expensive, high-latency round-trip across memory buses and PCIe interconnects.

What this means for AI training bottlenecks and data movement

The core bottleneck in large-scale AI training isn't always raw compute power — it's data movement. Shuttling training data between storage, DRAM, and GPU accelerators consumes significant energy and time. By offloading parameter calculations to the storage layer, Samsung's approach could reduce that overhead meaningfully, especially for workloads where the dataset is too large to fit in high-speed memory.

For the broader AI infrastructure market, this signals that Samsung — a dominant player in NAND flash and DRAM — is positioning its storage products as active participants in the AI compute stack, not just passive data containers. If this architecture makes it into real products, it could change the way data centers and edge AI systems are designed.

Editorial commentary on a publicly published patent application. Not legal advice.