KAIST unveils SoulMate on-device AI chip at ISSCC, targets 2027 commercialization
A KAIST research team unveiled a prototype AI semiconductor called SoulMate at the International Solid-State Circuits Conference (ISSCC) in the United States, showcasing on-device, highly personalized artificial intelligence that can learn a user’s speech style, preferences, and emotions. The team says the device processes all data inside the hardware, addressing privacy concerns, with productization planned around 2027 through a KAIST-affiliated startup named Onnyro AI.
SoulMate is described as an on-device accelerator for large language models that adapts to individual users. By running entirely within the device, it aims to reduce dependence on cloud servers, lowering latency and mitigating data-security risks associated with sending personal information over networks. The demonstration earned a “highlight paper” designation at ISSCC.
The researchers outline a hybrid approach that combines on-device retrieval-augmented generation with immediate on-device fine-tuning. A key feature is the Mixed-Rank architecture, which prioritizes important information during learning to speed up processing and cut energy use, while skipping repetitive training on similar sentences to save power.
In performance tests, the SoulMate chip reportedly responds to user input in about 216 milliseconds and achieves roughly five times higher power efficiency than typical mobile AI workloads. The chip operates at an ultra-low power level of 9.8 milliwatts while performing both learning and inference tasks, using Samsung’s 28-nanometer process to fit a 20.25 square-millimeter die.
KAIST’s lead researcher, Professor Yu, described on-device fine-tuning of a large language model as a world-first accomplishment, arguing that personalization now occurs within the device’s own processing core rather than through stored user data alone. The team emphasizes that the device’s “brain” is being updated, not just its memory of user information.
The development sits within a broader trend of increasing AI integration into smartphones, wearables, and home devices. Major tech players such as Samsung, Apple, and Google have been exploring or deploying on-device AI capabilities to deliver faster, privacy-preserving AI assistants without constant cloud access.
The project reflects KAIST’s position in hardware-focused AI research. The SoulMate prototype, highlighted at ISSCC, is intended as a path toward practical on-device personalization at scale, with commercialization targeted for around 2027 through Onnyro AI, a KAIST-founded startup.
For U.S. readers, the work matters because on-device AI promises lower network traffic, enhanced privacy, and faster responses for consumer devices, which could influence smartphone design, security standards, and supply chains. If scalable, such hardware could shift how AI features are delivered to American consumers, impacting device performance, energy use, and the competitive landscape for AI-enabled hardware.