Phison Electronics has unveiled a new AI memory extension approach called aiDAPTIV, designed to expand effective working memory for local artificial intelligence workloads by combining system DRAM with high-endurance NAND flash through its Pascari aiDAPTIV Cache Memory technology. The company says the solution helps reduce dependence on large DRAM configurations while enabling more complex AI models to run directly on consumer and enterprise PCs.
According to internal testing by Phison Electronics, the aiDAPTIV system allowed a 26-billion-parameter model to run on a machine with 16 GB of DRAM, whereas 32 GB would typically be required without the technology under the same conditions. The improvement is driven by techniques such as KV cache reuse and memory offloading, which extend effective AI working memory beyond physical DRAM limits.
The technology is being developed in collaboration with Intel and is optimized for AI PC platforms powered by Intel Core Ultra processors. It also integrates support for the OpenVINO toolkit, allowing developers and independent software vendors to evaluate and deploy AI workloads more efficiently across local devices.
Executives from both companies emphasized that AI PCs are evolving into systems capable of running increasingly complex agentic applications and mixture-of-experts models, which place heavier demands on memory and responsiveness. KS Pua, CEO and founder of Phison Electronics, said the goal is to enable OEMs, developers, and end users to run larger AI models locally while maintaining privacy and reducing infrastructure costs.
At Computex, Phison is expected to demonstrate aiDAPTIV-enabled systems running on Intel AI PC platforms. These demos will include a local chat interface powered by a MoE model that would normally exceed system memory limits, showing how the technology extends usable AI memory without requiring additional DRAM.
The company will also showcase a hybrid AI routing application built on OpenClaw, illustrating how local models can be combined with cloud services when necessary for more demanding tasks. Additional demonstrations will feature ecosystem partners including Ollama, LLMWare, TurinTech AI, as well as hardware collaborators such as ASUS, MSI, and Acer.
Industry partners involved in the ecosystem highlighted the importance of memory efficiency for local AI deployment. Ollama co-founder Michael Chiang noted that memory constraints remain a major barrier to running larger models on consumer hardware, while LLMWare emphasized the growing demand for on-device enterprise workflows such as retrieval-augmented generation and domain-specific agents. TurinTech AI added that combining AI optimization tools with improved memory architecture can help bring more practical AI workloads directly onto PCs.
Intel’s client computing leadership also stressed that users and businesses increasingly want faster, more private AI experiences without relying heavily on cloud infrastructure. The collaboration with Phison aims to enable larger models to run locally with simpler system configurations, improving both cost efficiency and data privacy.
Overall, aiDAPTIV positions itself as a bridge between hardware limitations and the growing demands of modern AI workloads, particularly as PCs transition into full-scale AI computing platforms.