AI Computing Done by Memory: Samsung Develops World's First LPDDR-PIM

[Financial News] Samsung Electronics has developed the world's first next-generation memory that can perform computing inside memory, targeting mobile and other on-device artificial intelligence (on-device AI) devices. The technology adds computing functions to mobile memory, which had previously only stored data, in order to speed up AI processing and reduce power consumption.
Hwang Garam, an engineer at Samsung Electronics, unveiled the structure, specifications and verification results of LPDDR5X-PIM, which combines low-power memory LPDDR with processing-in-memory (PIM) technology, at the Hot Chips 2026 semiconductor conference held at Leland Stanford Junior University (Stanford University) in the United States on the 25th local time. The product won a next-generation memory award earlier this month at the FMS 2026 memory conference, but this was the first time its detailed specifications were disclosed.
PIM is a technology that embeds computing units in memory so that some calculations can be handled directly. In the past, data stored in memory had to be sent to the central processing unit (CPU) or graphics processing unit (xPU) for computation and then brought back to memory. With PIM, relatively simple operations can be processed inside memory, reducing data movement, which is considered a major bottleneck for AI. That improves processing speed while also lowering power consumption.
In tests conducted by Samsung Electronics on Meta Platforms' open AI model Llama 3.1, which has 8 billion parameters, LPDDR5X-PIM completed processing in 5.4 seconds, about 2.3 times faster than the 12.3 seconds required by standard LPDDR5X. Token throughput per second (TPS) reached 81.3 TPS, three times higher than the 27 TPS of the previous product.
Samsung Electronics previously unveiled HBM-PIM in 2021, applying PIM technology to high-bandwidth memory (HBM), and carried out development verification. This time, it applied PIM to low-power LPDDR, which is mainly used in mobile devices such as smartphones, instead of HBM.
The move reflects the view that not every AI device needs HBM as AI models become smaller and more efficient. Hwang explained, "The standard memory solution for AI is HBM, but as large language models (LLMs) have recently become smaller and more efficient, HBM may not always be the best solution. That is why we developed LP-PIM, which combines high bandwidth and low power consumption."
The potential use of LPDDR-PIM is also growing as on-device AI, which runs generative AI directly on smartphones and personal computers (PCs) rather than in data centers, continues to spread. In mobile devices with limited power and space, it is important to improve AI computing performance while reducing power consumption.
Samsung also improved compatibility with existing systems. It designed the product to use the same array as general-purpose LPDDR, allowing it to be used without major system changes. Samsung Electronics is currently building the related ecosystem by providing a Software Development Kit (SDK), simulators and other tools.
Following LPDDR5X-PIM, Samsung Electronics has begun standardization work on "LPDDR6 PIM." As AI memory competition, which had centered on HBM, expands into on-device areas such as smartphones and PCs, memory is evolving beyond a device that simply stores data into one that performs computing directly.

[email protected] Kim Kyung-min Reporter