Floating Point Multiple-Precision Fused Multiply Add Architecture for Deep Learning Computation on Artix 7 FPGA Board
Deep learning (DL) has become a transformative force in today's world revolutionizing industries. However, its success relies on high-precision arithmetic units, leading to the requirement of powerful high precision arithmetic design. Hence, this research proposes the multiple precision fused...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Stefan cel Mare University of Suceava
2024-11-01
|
| Series: | Advances in Electrical and Computer Engineering |
| Subjects: | |
| Online Access: | http://dx.doi.org/10.4316/AECE.2024.04010 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Deep learning (DL) has become a transformative force in today's world revolutionizing industries. However,
its success relies on high-precision arithmetic units, leading to the requirement of powerful high precision
arithmetic design. Hence, this research proposes the multiple precision fused multiply add (MPFMA) architecture
for profound computing-based applications. The proposed MPFMA architecture is capable of performing momentous
tasks in every single clock cycle such as eight consecutive numbers of half precision (HP) operations, four
numbers of concurrent single precision (SP) operations, two simultaneous double precision (DP) operations
and single quadruple precision (QP) operations. The propounded architecture is implemented using Xilinx
Vivado 2022.2 on Artix-7 FPGA Basys 3 board that demonstrates the functionality and attainment. From the
observed results, it is inferred that the proposed framework offers 50% area curtail with the conventional
FMA architecture, while still meeting the precision requirements for deep learning tasks. With an astoundingly
low error rate of 0.013 % and an amazing accuracy rate of 99.987 %, the MPFMA in deep learning hardware not
only enhances model performance but also contributes to energy conservation, making DL systems more sustainable
and promising for the future of smart intelligence applications. |
|---|---|
| ISSN: | 1582-7445 1844-7600 |