-
Notifications
You must be signed in to change notification settings - Fork 13
/
Readme.txt
18 lines (14 loc) · 916 Bytes
/
Readme.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
Objective
A hardware coprocessor for CNN acceleration based on RISC-V instructions set, which is linked with Hummingbird E200 MCU through extension accelerator interface (EAI).
Hummingbird E200: https://github.com/SI-RISCV/e200_opensource
Features
• Support Convolution layer and ReLU layer
• Transform convolution into matrix multiplication (im2col on the fly)
• Kernel size from 3×3 to 11×11
• Support data type INT8,UINT8, EXP4 (4 bits of exponential scale) and Ternary
• 16 dot-product operations of 64-bit operands (8 INT8, 16 EXP4 or 32 Ternary) per cycle
• PE utilization is 100% for internal layers. For the input layers, the utilization is 82.5%(11×11×3),87.5%(7×7×3),93.75%(5×5×3),84.375%(3×3×3).
Please see ./spec/CNN-HWPE_SPEC_EN.pdf for more details
Author: Hao Chen , Qiang Chen
Email: chenhaocxjtu@163.com
chenqiang5233@hotmail.com