Recently, computing in memory (CiM) has been proven to be an energy-efficient and promising architecture for artificial intelligence (AI) algorithms. And yet, current CiM schemes generally suffer from limited throughput compared to their digital counterparts, and the key reason is that the CiM macro calculation must iterate through multiple cycles. Thus, the need to reduce the calculation cycle of the macro while keeping high energy efficiency and the necessity of developing acceleration methods for the universal CiM-based processor have become major issues faced by the current CiM architectures. To surmount these critical problems, we propose a processor based on a two-cycle CiM macro. Our work makes three main contributions: 1) we present a Radix16-based digital-CiM macro with look-up table (LUT) optimization to reduce dynamic power consumption; 2) we devise a hybrid Winograd microarchitecture and dataflow that supports (2, 3) and (4, 3) Winograd convolution, meaning that a good compromise can be reached between the accuracy of the algorithm and the reduction in workload; and 3) we propose a macrolevel parallel dual-side sparse CiM core that uses a horizontal direction compression method to reduce the input cycle of activation data and improve the mapping efficiency of the weight data in the macros. A prototype of the processor is fabricated in a 28-nm CMOS, which achieves a peak system energy efficiency of 19.9–258.5-TOPS/W for a voltage supply of 0.6–1.1 V, and an operating frequency of 78–287 MHz, a 2.55–7.08 $\times$ higher than other state-of-the-art CiM processors.
5 August 2025 in IEEE Journal of Solid-State Circuits
Computing-in-memory (CiM) architecture is considered a technology that promises to alleviate the memory wall problem in artificial intelligence (AI) computing. Previous work on digital CiM macros has achieved significant energy efficiency advantages in supporting approximate calculations of traditional integer (INT)/floating point (FP) data. POSIT is an emerging FP number system. Its main characteristic is that it can approach a good tradeoff between data range and accuracy by flexibly adjusting the mantissa bit width. However, current digital CiM macros lack the ability to support approximate calculations of POSIT data. Given this, we propose a POSIT-approximate-calculation-based CiM macro. The three main contributions of our work are as follows: 1) we improve the original POSIT approximate calculation principle and design a calculation logic circuit based on unsigned Radix8+ final-cycle fusion (FCF); 2) we propose a macro-level approximate calculation control strategy to help the CiM macro find a compromise between calculation performance and calculation error; and 3) we design an activation mantissa alignment skipping (AMAS) unit, to help the hardware skip sparse activation bits when calculating POSIT multiplication. A prototype of the CiM macro was fabricated in a 28-nm complementary metal oxide semiconductor (CMOS) and scored a peak macro energy efficiency of 88.3 TFLOPS/W @ POSIT (8, 2) for a voltage supply of 0.55–1.1 V and an operating frequency of 70–280 MHz, a value $2.12{\times }$ higher than other state-of-the-art POSIT CiM macros.
13 November 2023 in IEEE Transactions on Circuits and Systems I: Regular Papers
Super-resolution (SR) task using the convolutional neural network is a crucial task in improving image and video quality. The introduction of the residual block (RB) raises the depth of the algorithm to perform better reconstruction. The processing of the RB leads to a decrease in hardware utilization and frequent off-chip communications. It is hard to apply such algorithms on edge devices with limited performance. Computing-in-memory (CiM) is one promising method to reduce high power caused by massive data movement in multiply-accumulation computation. The algebraic sparsity (AS) is the structured sparsity (SS) optimization for imaging computing. However, it is an unsolved problem to simultaneously realize the texture sparsity (TS) of the image and the SS of the algorithm in the CiM scheme while maintaining high hardware utilization. Thus, we propose a CiM-based SR task accelerator. There are three key contributions: first, a texture-aware workflow and a dynamic grouping CiM engine can concurrently support TS coupling with AS. Second, a macro-level pipeline scheme together with two custom-sized CiM macros and a high reuse-rate Hadamard transformation circuit reaches 91% hardware utilization. Third, a novel weight update strategy is devised to reduce the performance loss induced by the weight updating. The accelerator prototype is fabricated in a 28-nm CMOS. It scores a 22.8-44.3-TOPS/W peak energy efficiency at the voltage supply of 0.54-1.1 V and the operating frequency of 50-200 MHz, indicating 1.8-6.8x higher compared to the state-of-the-art CiM processors.
7 April 2026 in IEEE Journal of Solid-State Circuits
Static random-access memory (SRAM)-based computing-in-memory (CIM) macros have been widely studied to improve the energy efficiency of edge artificial intelligence (AI) inference tasks. However, less attention has been given to AI training, which requires CIM macros to not only perform matrix multiply-accumulate (MAC) operations but also support matrix transposition. To address the limitations of previous analog transpose and digital non-transpose SRAM CIM macros, this work features: 1) a cyclic-weight-mapping SRAM array that enables matrix transposition and reuse of MAC circuits during both feed-forward (FF) and back-propagation (BP) phases; 2) a digital CIM architecture employing signed fixed-point mantissa encode and a vector-wise pre-alignment (VWPA) scheme, supporting multiple data formats including INT4/8, FP8, and BF16; and 3) an accurate/approximate dual-mode bit-parallel MAC circuit (DMBP-MAC) designed to provide a tradeoff between computational accuracy and energy efficiency. A fabricated 28-nm 32-kB transpose SRAM CIM macro achieved average energy efficiency of 70.2–285.4 TOPS/W in INT4, 17.5–71.4 TOPS/W in INT8, 51.1–192.3 TFLOPS/W in FP8, and 12.8–48 TFLOPS/W in BF16.
3 March 2026 in IEEE Journal of Solid-State Circuits
With increasingly sophisticated artificial intelligence (AI) applications, there is an escalating need for computing-in-memory (CIM) schemes to facilitate high-precision floating-point (FP) multiply accumulate (MAC) computing, while existing digital or analog CIM implementations exhibit inherent limitations when confronted with high-bit precision situations. Therefore, this work presents a hybrid-domain CIM macro to enhance the energy efficiency of FP CIM computation. This work employs: 1) a hybrid-domain CIM macro structure that fully utilizes the advantages of digital and analog CIM paradigms; 2) a logarithm bit-width zero-canceling residual analog-digital convertor (ADC) to minimize the ADC overhead and improve throughput; and 3) an outer product-based FP/INT CIM architecture to eliminate the need for a large fan-in multi-level adder tree and reduce the transmission delay. A fabricated 28-nm 192-Kb INT/FP CIM macro achieves maximum peak energy efficiency of 72.12 TFLOPS/W and maximum area efficiency of 2.21 TFLOPS/mm2. The accuracy loss of running ResNet-18 on the CIFAR-100 dataset with BF16IN-BF16W-BF16OUT is 0.05%.
10 November 2025 in IEEE Journal of Solid-State Circuits
Convolutional neural network (CNN) and transformer are the most popular neural network models in computer vision (CV) and natural language processing (NLP). It is quite common to use both these two models in multimodal scenarios, such as text-to-image generation. However, these two models have very different memory mappings, dataflows and mathematical operators, making it difficult to accelerate both types of models simultaneously. To address the forementioned challenges, we propose a multi-core programmable near-memory accelerator and introduce an arbitration-free multi-port static random-access memory (SRAM) array to improve storage utilization while maintaining flexibility. To achieve performance comparable to computing-in-memory (CIM) designs, we use near-memory variable-precision multiplier-accumulators (NVMACs) to perform multiply-accumulate (MAC) operations tightly close to the memory to maximize the memory access throughput and support the mixed-precision neural network inference. We use a fine-grained instruction set architecture (ISA) to support software sparsity and reduce overhead caused by coarse-grained non-MAC operations with low utilization. A chip is fabricated in a 28 nm process and achieves 6.3-to-101.4TOPS/W energy efficiency for transformer model and 7.3-to-194.6 TOPS/W for CNN model, $1.2\times $ to $4.2\times $ compared with other state-of-the-art designs, while efficiently supporting both CNN and transformer overloads.