Server-Hardware

Samsung Brings Compute Into Memory: LPDDR5X-PIM Triples AI Speed

At Hot Chips 2026, Samsung unveiled its world-first processing-in-memory for low-power DRAM: LPDDR5X-PIM is said to roughly triple AI inference compared to conventional LPDDR5X – without device makers having to redesign their boards.

Computing right inside memory

PIM stands for “Processing-in-Memory”: a small logic unit sits right next to the DRAM cells and performs simple calculations locally. This bypasses the processor as a bottleneck and speeds up AI inference in particular. Samsung cites 3.01x AI throughput and 614 GB/s PIM bandwidth – in the same 561-ball package makers already use, so no PCB redesign is needed.

A cheaper alternative to HBM

Expensive, power-hungry HBM memory sits inside data-center AI accelerators. LPDDR5X-PIM targets the tier below: smartphones, AI PCs and edge devices, where HBM is too costly and complex. Industry observers see it as a possible antidote to “chipflation” – the price pressure from the concentration of scarce memory chips.


Sources: Tom’s Hardware, ServeTheHome.

Mastodon
Scroll to Top