Tri-Oracle: A 17.78μJ/Token Vision-Language Model Accelerator with Token-Attention-Weight Redundancy Prediction
Abstract
This paper presents Tri-Oracle, a VLM accelerator exploiting token, attention, and weight redundancies. A Token Merging Unit (TMU) merges 68% of redundant tokens. An Attention Head Prediction Unit (AHPU) predicts streaming heads and skips them, reducing attention latency by 44% and balancing load. A Strip-wise SNZV Computation Unit (SSCU) predicts near-zero values, cutting FFN latency by 45% and EMA by 57%. Fabricated in 28 nm CMOS, Tri-Oracle achieves$17.78 \mu \mathrm{J} /$token and 1.19 to$3.99 \text{TOPS} / \text{mm}^{2}$without fine tuning.
Fields of Study & Subjects
Advanced Neural Network Applications
Low-power high-performance VLSI design
Advanced Memory and Neural Computing
Publication details
- Journal / source
- Digest of technical papers/Digest of technical papers - IEEE International Solid-State Circuits Conference
- Pages
- 542–544
- Publication date
- 1 February 2026
- Language
- EN