– Title: A 333TOPS/W Logic-Compatible Multi-Level Embedded Flash Compute-In-Memory Macro with Dual-Slope Computation
– Conference: 2023 IEEE Custom Integrated Circuits Conference (CICC)
– Abstract: Computing-in-memory (CIM) has been an ongoing prominent research area for easing the energy efficiency of machine learning tasks in edge devices. Recently, embedded non-volatile memory (eNVM) CIM architectures have been popular as an edge device, where it can turn off their supply during standby for low power consumption. However, most eNVMs (e.g., MRAMs and RRAMs) require the use of specialized technologies and are mostly used as single-level cell (SLC) data storage. In the technologies that do not provide eNVMs, logic-compatible single-poly non-volatile embedded flash (eflash) memory can be considered an alternative. Although the cell area of the single-poly non-volatile eflash is significant, we can considerably compensate for the cell area penalty by using multi-level cells (MLCS). Further, in eNVM CIMs, the analog computations must be quantized with an ADC, where the SAR ADCs are a popular conversion topology. However, SAR ADC designs result in significant power consumption and area overhead due to its capacitor DAC driving and high accuracy comparators. In this work, we propose to overcome such challenges by proposing 1) a logic-compatible single-poly nonvolatile eflash memory macro using MLC and an SLC at the same time to increase computation density while maintaining a reasonable signal margin, 2) a resolution configurable differential SAR TDC used for both memory programming and computing with replacing analog voltage comparators to inverters in order to reduce power consumption and area, and 3) an energy-efficient 2’s complement dual-slope computation with MLC and SLC sharing a single differential TDC for multi-bit weight computation. We fabricated the proposed eFlash CIM macro in a 65 nm CMOS process. Our measurements show that the proposed CIM macro achieves up to 333 TOPS/W energy efficiency and 186.2 GOPS throughput.
– Main figure
