How Near‑Memory Computing Can Power Edge LLMs: A 2025 Storage Framework

Large language models (LLMs) have transformed natural language processing since the 2017 introduction of the Transformer architecture, with models like OpenAI's GPT series and Meta's LLaMA gaining prominence. Traditionally, these models run on cloud servers, which provides ample compute power but introduces drawbacks such as network latency, data‑security concerns, and the need for continuous connectivity, limiting real‑time user experiences.

To address these issues, a new computing paradigm—integrated storage and computation—has emerged. By embedding compute capabilities directly within memory devices, this approach can dramatically improve data‑movement efficiency and overall energy efficiency, overcoming the von Neumann bottleneck. The integrated architecture can be categorized into three main types:

Near‑Memory Computing (PNM): Places compute units close to memory modules, shortening data paths, increasing memory bandwidth, and boosting performance for workloads that require massive parallelism and optimized bandwidth.

Processing‑In‑Memory (PIM): Embeds compute units inside the storage chip itself, granting the memory intrinsic processing ability. This is ideal for data‑intensive tasks, offering significant gains in processing efficiency and energy‑performance ratio.

Computing‑In‑Memory (CIM): Deeply fuses storage and compute, allowing memory cells to directly participate in data processing. CIM excels in highly parallel computations and custom hardware optimizations, effectively eliminating memory‑access latency. The choice among PNM, PIM, and CIM depends on specific application requirements and performance goals.

These storage‑centric solutions are especially relevant for the commercialization of edge‑side AI models, where minimizing latency and power consumption is critical. The article concludes that selecting the appropriate near‑memory technology is a strategic decision driven by the target workload, desired efficiency, and hardware constraints.

For a complete analysis and detailed framework, refer to the full report titled “Computer Industry Research: Edge‑Side Large‑Model Near‑Memory Computing and Customized Storage Research Framework (2025)”.