Samsung has unveiled LPDDR5X-PIM, a new memory design that moves AI computation closer to the data itself by placing processing logic directly alongside DRAM cells, detailed at the Hot Chips 2026 conference and reported on August 26, 2026. The technology represents Samsung's latest effort to address one of the most persistent bottlenecks in AI inference performance: the time and energy cost of repeatedly moving data between memory and separate processing units.
Processing-in-memory architecture allows certain calculations to occur inside the memory chip itself rather than requiring data to be shuttled back and forth to a separate processor, an approach Samsung says can dramatically reduce a major bottleneck affecting AI inference workloads. In preliminary testing, the company reported that LPDDR5X-PIM delivered 2.28 times faster model runtime and 3.01 times greater token throughput compared with conventional LPDDR5X memory, metrics that, if validated in real-world deployment, would represent a substantial efficiency improvement for AI systems constrained by memory bandwidth rather than raw computational power.
Perhaps most striking among Samsung's reported figures is the bandwidth improvement: peak bandwidth reportedly rose from 76.8 gigabytes per second on standard LPDDR5X-9600 memory to a theoretical 614 gigabytes per second in PIM mode, representing an eightfold increase. Samsung tested the design using an 8-billion-parameter Llama 3.1 model running on an edge AI accelerator, a configuration chosen to reflect the kind of on-device or edge-deployed AI inference scenarios where memory bandwidth constraints are often particularly acute given the more limited power and thermal budgets available compared with data centre environments.
The company has acknowledged that optimisation work, including accuracy tuning, remains ongoing, suggesting the benchmark results, while promising, should be treated as early-stage findings rather than fully finalised production performance figures. Nonetheless, the technology's unveiling at Hot Chips — one of the semiconductor industry's most closely watched annual conferences for new chip architecture disclosures — signals Samsung's intent to position itself as a leading innovator in memory technology specifically tailored to the demands of AI inference, a segment of the broader AI hardware market that has grown enormously in importance as AI products move from research and training environments into widespread commercial deployment.
The development arrives amid intensifying competition across the memory chip industry to address AI-specific bottlenecks, as the explosive growth in AI inference workloads has placed unprecedented demand on memory bandwidth and capacity, contributing to broader supply constraints and price increases across the memory market. Nvidia has separately warned enterprise customers of AI chip price increases exceeding 15 percent driven by rising memory costs, underscoring how memory technology has become a critical, and increasingly costly, chokepoint within the broader AI hardware supply chain.
For edge AI applications specifically — AI models deployed directly on devices rather than relying on cloud-based inference — technologies such as LPDDR5X-PIM could prove particularly consequential, given the tighter power and thermal constraints that edge devices face compared with data centre AI accelerators. Improved memory bandwidth efficiency at the edge could enable more sophisticated AI models to run directly on smartphones, wearables and other resource-constrained devices without requiring constant connectivity to cloud-based processing.
As Samsung continues refining LPDDR5X-PIM toward production readiness, the technology's ultimate commercial impact will depend on how quickly it can be integrated into the broader ecosystem of AI accelerator designs and edge device manufacturers, and whether the efficiency gains demonstrated in Samsung's preliminary testing translate consistently across the diverse range of AI model architectures and workloads that the technology would need to support at commercial scale.
