Vellex Computing

 

Vellex Computing 是一家源自史丹佛大學，並獲美國國家科學基金會 (NSF) 與能源部 (DOE) 支 持的類比半導體公司，致力於大幅提升 AI 訓練的能源效率。現今的 AI 模型訓練往往需要耗費百 萬瓦的電力與龐大的雲端運算成本，讓多數邊緣應用難以負擔。我們透過類比電路與類比運算架 構而非傳統的數位暴力運算，解決了 AI 訓練核心的最佳化問題。這項專利架構在毫瓦級超低功 耗下，實現了高達 17,000 倍的驚人運算加速。從能源、工業物聯網、機器人到遙測等邊緣 AI 應 用切入，Vellex 讓終端設備能在無須依賴雲端的情況下，安全地進行即時與持續性學習。憑藉著 多項商業測試合作與不斷成長的專利組合，Vellex 正引領下一代無雲端依賴的智慧邊緣應用。

 

技術亮點

• Physics-Based Analog Compute: Maps AI optimization problems directly onto analog circuits, which settle to a solution through physical dynamics rather than iterating through discrete digital steps. The result is an AI training engine that operates within the power budget of embedded and battery-powered systems. 
• Zero Data Movement: Computes in-situ where data resides, eliminating the repeated transfers between memory and compute units that dominate power consumption in conventional digital training. 
• Energy Efficiency and Speed: Delivers up to 17,000x processing speedup over conventional approaches while operating at milliwatt power levels. 
• On-Device Continuous Training: Enables battery-operated devices, including autonomous drones, satellites, IoT sensors, and humanoid robots, to train continuously on live data without cloud connectivity. Models adapt in real time, eliminating the cost and latency of cloud retraining pipelines.
• On-Device Data Security: Field data stays on the device. No transmission to external servers means no exposure to network-based risks and no dependency on internet connectivity.