top of page

N series: Model weight/Activation for AI

N-series IPs offer an efficient, lossless solution for reducing the storage and bandwidth demands of AI models. By compressing both model weights and activations, it significantly lowers data traffic power consumption, cache SRAM cost, and DRAM space usage. The algorithm achieves near-theoretical compression ratios and maintains consistent performance across different models. With minimal hardware cost, ultra-low latency, and high throughput, the solution features an adaptive, entropy-aligned design and a parallel hardware architecture that scales to meet mainstream DRAM bandwidth requirements. 

圖片 1.png

Note :

If specifically for CNN, Activation also can be described as ‘Feature Map’.

◇◆TITC N-Series IP

nseries_spec.png
nseries_1.png
nseries_2.png

Tel: +886-3-5829011 ||| Fax: +886-3-5829028

Rm. 308, Bldg.52, No.195, Sec 4, Chung-Hsing Rd., Jhudong, HsinChu, 310, TAIWAN

Copyright © TITC. All Rights Reserved

bottom of page