Phison is extending its aiDAPTIV+ SSD technology to target AI key-value cache workloads, aiming to expand effective GPU memory and speed up inference on personal computers.
The SSD controller and drive company announced aiDAPTIV+ in August 2024 as a way of using Phison’s Pascari SSD in small-scale AI training work with a proprietary LLM. Storage supplier StorONE is using aiDAPTIV+ for its ONEai automated AI system for enterprise storage. Now Phison has extended the product to expand GPU memory and accelerate AI inference, which it says will “significantly increase memory capacity and simplify deployment to unlock large-model AI capabilities on notebook PCs, desktop PCs, and mini-PCs.”
Michael Wu, President and GM of Phison US, said: “As AI models grow into tens and hundreds of billions of parameters, the industry keeps hitting the same wall with GPU memory limitations. By expanding GPU memory with high-capacity, flash-based architecture in aiDAPTIV+, we offer everyone, from consumers and SMBs to large enterprises, the ability to train and run large-scale models on affordable hardware. In effect, we are turning everyday devices into supercomputers.”
This is virtually the same message that Nvidia CEO Jensen Huang put out yesterday at his CES 2026 presentation in which he talked about context memory extension to NVMe SSDs controlled by the BlueField-4 DPU.
But Huang was talking about rack-scale Vera Rubin pod inference workloads whereas Wu is thinking more about notebook PCs, desktop PCs and mini-PCs – the other end of the edge-to-datacenter spectrum.
Combining Phison’s aiDAPTIV+ and new Intel Core Ultra Series 3 processors, with built-in Intel Arc GPUs, enables larger LLMs to be trained directly on notebook PCs. Phison’s lab testing shows that a notebook equipped with this technology can fine-tune a 70B-parameter model.
Early aiDAPTIV+ inference testing on notebook PCs shows substantial responsiveness gains, delivering improvements in Time to First Token (TTFT).
Phison says that, for Mixture of Experts (MoE) inference processing, aiDAPTIV+ offloads the memory demands from DRAM over to cost-effective, flash-based, cache memory; a Phison Pascari SSD for example. In Phison testing, a 120B-parameter model can now be handled with 32 GB of DRAM in contrast to the 96 GB required in traditional approaches. This expands the ability to do MoE processing to a broader range of platforms.
Phison is working with Acer to unlock larger LLM training with Acer laptops using less DRAM resources. This enables users, the two say, to run AI workloads on smaller platforms with the required data privacy, scalability, affordability and ease of use. Mark Yang, AVP, Compute Software Technology at Acer, said: “Our engineering collaboration enables Phison’s aiDAPTIV+ technology to accommodate and accelerate large models such as gpt-oss-120b on an Acer laptop with just 32 GB of memory.”
