On-desk, Faster & Cheaper AI Training with Vellex Computing
Vellex Computing is a US-based startup focusing on developing high-speed, low-power edge computers for personalized & learning AI, delivering GPU-level performance within a hearing-aid energy budget. As one of the winners of the IC Taiwan Grand Challenge Batch 4 in the AI Core Technologies and Chips category, Vellex Computing specializes in programmable analog circuits for optimization with sub-millisecond latency and milliwatt power.
Resolving the Edge AI Trade-off
Current edge AI is trapped in a trade-off between intelligence and energy. Utilizing high-performance digital accelerators are capable of both training and inference, but will consume 5W to 10W, which is enough to drain a drone or hearing aid battery in minutes. On the other hand, ultra-low-power "wake-up" solutions are limited to static sensing and will be limited to inference only. This leaves today's edge devices "static"—incapable of learning locally without relying on the expensive, high-latency cloud. In addition, training a basic AI model costs more than USD 100,000 in cloud credits; which is far more expensive than what startups and independent researchers can afford.
Vellex Computing transforms AI from a power-hungry digital process into a high-speed physical event, transitioning Static Inference to Dynamic Learning-at-the-Source. By solving optimization needed for AI training using programmable analog circuits, Vellex Computing enables edge devices to learn and adapt in real-time within a milliwatt power budget.
Introducing IC Taiwan Grand Challenge
To strengthen Taiwan's position as a global semiconductor powerhouse, the National Science and Technology Council (NSTC) organizes the IC Taiwan Grand Challenge. Started in 2024, the competition originally focused purely on innovative IC designs and advanced application solutions. Today, ICTGC also honors teams in the fields of AI Core Technologies and Chips, Smart Mobility, Smart Manufacturing, Smart Medtech, and Sustainability, which can support and contribute to the semiconductor industry and Taiwan’s industry in general. Vellex Computing was selected as one of the winners of ICTGC Batch 4 in the AI Core Technologies and Chips category by enabling the only architecture purpose-built for affordable, on-device AI training.
Spearheading an AI Paradigm Shift and Democratizing High-Performance AI
Vellex Computing was founded in 2021 and is currently in the strategic pre-commercialization phase, focusing on silicon validation through critical collaborations with Tier-1 industry leaders. Vellex Computing provides a programmable analog co-processor IP designed to work alongside standard MCU cores, such as Arm or RISC-V, to handle the most compute-intensive tasks at the source. This specialized IP enables high-speed model training directly on the edge device, allowing for rapid weight updates in milliseconds with ultra-low power consumption. Vellex Computing transforms AI from a power-hungry digital process into a high-speed physical event. They utilize the natural physical dynamics of circuits to perform mathematical optimization. The solution is not "calculated" in the digital sense; rather the circuit is allowed to "settle" into its lowest energy state, allowing electrical currents to converge to the mathematical minimum through natural physical settling and to eliminate the need for the massive electrical current required to drive traditional digital iterations. As a result, Vellex Computing has achieved a validated 17,000x speedup, transforming a 2-hour digital training workload into a 390-millisecond physical event, while operating entirely within a milliwatt power budget. By solving optimization needed for AI training using programmable analog circuits, they enable edge devices to learn and adapt in real-time within a milliwatt power budget.