Samsung Patents a Method to Store AI Reference Data Directly on a Dedicated Drive
Training large AI recommendation models requires enormous lookup tables that regularly overwhelm system memory. Samsung's new patent shifts those tables onto a specialized drive that can talk directly to the GPU, skipping the usual memory bottleneck.
How Samsung's semantic SSD handles AI lookup tables
A warehouse worker has to fetch a specific box from a massive shelf before anyone can do anything useful with it. In AI systems that power recommendation engines (think: what movie to suggest next, or which ad to show), the model constantly does something similar, fetching numeric codes called embedding vectors from a giant lookup table before it can learn anything.
The problem is that these tables can be enormous, far too big to fit comfortably in the fast memory a GPU typically uses. Samsung's patent proposes storing the table on a specialized solid-state drive called a memory semantic SSD, which can hand vectors directly to the GPU and receive updated ones back after each round of training, without clogging up the rest of the system.
The goal is to make AI training faster and cheaper by letting the drive handle storage while the GPU focuses on number-crunching. You'd likely never see this hardware yourself, but it would sit inside the data centers that run the AI behind your daily recommendations.
… looking up embedding vectors corresponding to the input data from an embedding table of a model stored in a memory semantic solid state drive (MS SSD); transferring the embedding vectors acquired from the MS SSD to a graphics processing unit (GPU); and instructing the GPU to transfer updated embedding vectors to the MS SSD for updating the embedding table …
Translation: The system pulls AI data straight from a specialized drive, sends it to the graphics card, and saves the newly trained results back.
How the MS SSD and GPU split the embedding workload
At the center of this patent is a component called a memory semantic SSD (MS SSD), a solid-state drive designed to support fine-grained, direct reads and writes from a GPU rather than routing everything through a CPU or main memory.
The process works in a loop:
- The system receives input data (for example, a user ID and a list of items they interacted with).
- It queries the embedding table stored on the MS SSD, fetching the numeric vectors that represent those inputs. Embedding vectors are essentially compressed descriptions of things like users, products, or words, encoded as rows of numbers the model can do math on.
- Those vectors move to the GPU, which uses them in a training step, updating the numbers to make the model more accurate.
- The GPU then writes the updated vectors back to the MS SSD, so the table reflects the latest state of the model.
The key claim is that the MS SSD sits in the data path between storage and GPU in a way that avoids the conventional bottleneck of loading massive tables into CPU-managed RAM first. The patent covers the method, the apparatus managing the lookup-and-update cycle, and the system as a whole.
What this means for large-scale AI training costs
Recommendation AI is one of the most compute-intensive workloads in the world. Companies like Meta have published that their embedding tables can run to terabytes, sizes that simply don't fit in GPU memory and require expensive workarounds. A drive that can handle those tables natively and communicate with the GPU directly could cut both the hardware footprint and the time each training cycle takes.
Samsung's steady investment in storage-meets-AI filings fits a clear pattern: the company makes both the memory chips and the SSDs, so it has a direct incentive to expand what those components can do. Whether this specific architecture reaches production depends on how much it actually reduces latency in real workloads, something the patent doesn't measure.
This is the 50th Samsung filing we've tracked in the AI chip wars since June, building on ideas like a two-stage reading method and collapsing steps onto one chip.
Moving the lookup dictionary a model consults millions of times during training off the expensive processing card and onto a specialized storage drive frees up room for actual computation. The cost of that trade is speed: every training step now depends on how fast that drive can read and write, and any lag there stalls the whole process.
The setup also demands specific hardware talking to other specific hardware with very little friction, which makes it poorly suited to the mixed, already-purchased equipment most organizations are running.
For organizations training models so large that the lookup dictionary alone overwhelms what any single machine can hold, this is a reasonable answer to a real problem. Whether it stays reasonable depends on how easily Samsung can fit this into existing facilities without requiring a full hardware overhaul first.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
19 drawing sheets from US 2026/0267843 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →