Overview
- Phison introduced aiDAPTIV at Computex 2026 in Taipei and demonstrated running a 26-billion-parameter model on a system with 16 GB of RAM by using Pascari AI100E SSDs as an AI cache for a local chat and hybrid cloud workflow.
- The technology extends working memory by storing model key-value data and session tokens on high-performance NAND flash and dynamically swapping those items with DRAM to free RAM for model execution.
- Runtime features such as KV cache reuse are designed to cut repeated computation and speed time-to-first-token, improving responsiveness as context windows grow.
- The effort is positioned around Intel Core Ultra platforms with OpenVINO tooling to help OEMs and developers optimize apps, but it depends on manufacturers shipping Pascari or equivalent drives.
- Adoption questions remain because the solution relies on specialized high-end SSDs that are costly — a 1 TB Pascari AI100E appeared at $2,516 in a retailer listing — and Phison’s claims come from demos that need independent benchmarks and testing of endurance, power and software support.