Return
Calibration-Free Edge-Computing IMC Macro With Direct Current-to-Digital Conversion
A
R
L
A
G
DOI:10.1109/TCSII.2025.3581384.png)
Abstract
En 中文
In-memory computing (IMC) has emerged as a promising solution to the “memory wall” problem of traditional Von Neumann architectures by integrating computation directly within the memory. This brief presents a novel current-mode IMC macro that leverages a nanoampere-range temperature-independent reference current and direct current-to-digital conversion. The proposed design mitigates power inefficiencies and thermal instability of previous architectures without the need for calibration. Implemented in TSMC LP 65 nm CMOS technology, the design achieves an energy efficiency of 310.7 TOPS/W and an area efficiency of 18.93 TOPS/mm2. The reference current generator ensures a temperature coefficient of just 363 ppm/°C over a temperature <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\in $ </tex-math></inline-formula> [-10; 115]° C. Using a VGG-6 model on the CIFAR-10 dataset with 87.93% accuracy, the drop between the software baseline model and IMC hardware (1bA/1bW/5bO) at TT@27°C/FF@115°C/SS@-10°C is 0.17/0.32/0.51%.
Keywords:
Calibration-free
current source
edge computing
in-memory computing
multiply-accumulate
thermal stability
ultra-low power
Journal
I
IF:
4.9
Papers:
8.8K
Citations:
2.5W
