Samsung Electronics has introduced the world's first next-generation solution that performs computations directly within memory, targeting on-device artificial intelligence (AI) devices such as mobile phones.

Hwang Garam, a principal engineer at Samsung Electronics, unveiled the architecture, specifications, and verification results of 'LPDDR5X-PIM'—a commercialized Processing-in-Memory (PIM) product based on low-power LPDDR memory—at the 'Hot Chips 2026' semiconductor symposium held at Stanford University on the 25th local time.

While the product first made its presence known to the industry by winning the 'Next-Gen Memory Award' at the 'FMS 2026' memory conference earlier this month, this marks the first time its detailed specifications have been made public.

Unlike conventional architectures where data is moved from memory to processors such as central processing units (CPUs) or graphics processing units (GPUs) for calculation and then transmitted back to memory, PIM allows simple computations to be handled directly by arithmetic logic units built inside the memory.

As a result, processors only need to handle complex calculations, which enhances overall efficiency. By minimizing data transmission between the memory and the processor—traditionally a bottleneck in AI computing—it saves both time and power consumption.

When Samsung tested Meta's lightweight open-source model 'Llama 3.1' with an 8-billion parameter scale, the processing time using LPDDR5X-PIM was 5.4 seconds, about 2.3 times faster than the 12.3 seconds of conventional standard LPDDR5X.

The token processing speed per second (TPS) also reached 81.3 TPS, nearly triple the conventional 27 TPS.

Samsung previously introduced 'HBM-PIM' by applying similar PIM technology to High Bandwidth Memory in 2021 and completed development validation, but this commercialized product was implemented by applying it to LPDDR instead of HBM.

Principal Engineer Hwang explained the background of the development, stating, "While High Bandwidth Memory (HBM) is the standard memory solution supporting AI, HBM may not always be the best solution as large language models (LLMs) recently become smaller and more efficient. Consequently, we developed LP-PIM, which delivers both high bandwidth and low power consumption."

Because the layout is identical to existing general-purpose standard LPDDR specifications, compatibility is high, allowing this product to be installed without modifying existing systems, according to the explanation.

As the performance of on-device AI emerges as a crucial factor in choosing mobile devices, Hwang expects that LPDDR memory equipped with computing functions will drive solid demand in the market.

Hwang added that the company is currently building an ecosystem by providing a software development kit (SDK) and simulator, and has also begun standardization work for the next-generation LPDDR6-PIM.

(Photo: Yonhap News)