Samsung Electronics has commercialized the world's first technology that performs artificial intelligence computations directly within memory. As demand for High Bandwidth Memory (HBM) surges centered around AI data centers, Samsung has introduced a novel memory solution tailored for on-device AI by adding computing capabilities to LPDDR, a low-power memory primarily used in smartphones and PCs. This breakthrough dramatically reduces data movement, boosting both processing speed and power efficiency.
Key Performance and Technological Breakthroughs
At the 'Hot Chips 2026' semiconductor academic conference held at Stanford University, Samsung Electronics Researcher Garam Hwang unveiled the architecture, specifications, and performance verification results of 'LPDDR5X-PIM,' a Processing-in-Memory commercial product based on LPDDR.
Traditionally, data stored in memory had to be transferred to a Central Processing Unit (CPU) or Graphics Processing Unit (GPU) for computation, and the results sent back to memory. PIM architecture, however, enables simple computations to be processed directly inside the memory chip.
In actual performance tests using Meta's 8.0-billion-parameter lightweight open-source AI model, 'Llama 3.1,' the LPDDR5X-PIM demonstrated remarkable efficiency:
Processing time was slashed to 5.4 seconds, roughly 2.3 times faster than the 12.3 seconds recorded by standard LPDDR5X.
Token Processing Speed (TPS)—representing the number of tokens processed per second as the AI comprehends sentences and generates responses—surged from 27 TPS to 81.3 TPS, a roughly threefold increase.
Earlier this month, Samsung also showcased LPDDR5X-PIM at the 'FMS 2026' memory conference, where it was honored with the Next-Generation Memory Award.
Expanding AI Memory Beyond Data Centers to Edge Devices
Samsung’s integration of PIM into LPDDR is a strategic move to expand the application scope of the AI memory market from data centers to mobile and edge devices.
While HBM remains the core memory for massive data processing in AI data centers due to its high bandwidth, it poses significant challenges regarding power consumption and packaging overhead. On the other hand, LPDDR excels in low power consumption, making it ideal for smartphones and laptops. By combining LPDDR with PIM, Samsung has successfully secured both the high performance and low power efficiency required for on-device AI.
"While HBM is the standard memory solution supporting AI, as Large Language Models (LLMs) continue to shrink and become more efficient, HBM may not always be the optimal solution," explained researcher Hwang regarding the development background. "Consequently, we developed LP-PIM to simultaneously deliver high bandwidth and low power consumption."
Seamless Compatibility and Future Roadmap
A standout feature of the LPDDR5X-PIM is its design, which utilizes the exact same memory array as conventional LPDDR to maximize compatibility. Samsung explained that this allows manufacturers to adopt the new product without heavily altering existing system architectures, thereby reducing the burden of system redesign and boosting commercial adoption potential.
Following its introduction of 'HBM-PIM'—applying PIM technology to HBM in 2021—Samsung has now successfully broadened its application scope to LPDDR. The company currently provides software development kits (SDKs) and simulators to foster a robust ecosystem and has already initiated standardization work for the next-generation LPDDR6-PIM.
Industry experts note that Samsung's commercialization of LPDDR5X-PIM holds profound significance as it introduces a fresh alternative to an AI memory market heavily saturated by HBM. As AI expands beyond centralized data centers into consumer devices like smartphones and PCs, the demand for low-power, high-efficiency memory optimized for specific device characteristics is expected to grow exponentially.
[Copyright (c) Global Economic Times. All Rights Reserved.]
