| IP개요 |
It presents EVO-VITA, a versatile and energy-efficient AI processor, designed to realize on-device agentic AI systems with both inference and training capabilities. Unlike conventional NPUs, which were limited to inference, EVO-VITA supports back-propagation, low-rank adaptation, and backward-unlocking-based training through numerous features using hardware-software co-optimization. The processor incorporates a novel inference and training unified core for hybrid-data-representation-based main convolution or matrix multiplication acceleration, a unified peripheral core for specialized tensor operations, and a memory management unit for efficient off-chip data handling. The 4 mm × 4 mm chip targets highest throughput and highest energy efficiency compared with the conventional state-of-the-art inference or training processors. Moreover, the fabricated chip will demonstrate two scalable applications, from AIoT devices to edge base stations, using single-chip and multi-chip boards, respectively |
