RRAM is a promising candidate to implement large-capacity in-memory computing on edge AI devices due to its high density. However, the efficiency and accuracy of RRAM-based computing-in-memory (CIM) works are limited by large accumulation currents and device variations. The SRAM-based digital CIM achieves superior performance and efficiency while the adder tree dominates the area. In this paper, a digital RRAM CIM macro is proposed to achieve a better trade-off between accuracy, energy, and performance by three techniques. First, a dynamic voltage sense amplifier is designed to reduce >90% read currents of low resistance state (LRS) cell. Second, an OR gate approximate adder tree is proposed to reduce the area of the adder tree by 40%. Third, a sparse-aware finetuning algorithm is proposed to reduce the accuracy loss of approximate arithmetic to 0.4% on Cifar-10 dataset. The proposed design achieves 1966TOPS/W energy efficiency and 51.2GOPS/Kb normalized throughput at 1-bit precision which is 1.2× and 1.8× higher than previous RRAM-based designs. This work demonstrates the advantage of digital CIM using RRAM devices.
30 October 2023 in IEEE Journal of Solid-State Circuits
Computing-in-memory (CIM) chips have demonstrated the potential high energy efficiency for low-power neural network (NN) processors. Even with energy-efficient CIM macros, the existing system-level CIM chips still lack deep exploration on sparsity and large models, which prevents a higher system energy efficiency. This work presents a CIM NN processor with more sufficient support of sparsity and higher utilization rate. Three key innovations are proposed. First, a set-associate blockwise sparsity strategy is designed, which simultaneously saves execution time, power, and storage space. Second, a ping-pong weight update mechanism is proposed for a higher utilization rate, enabling simultaneous execution of CIM and write operations. Third, an efficient CIM macro is implemented with adaptive analog-digital converter (ADC) precision for better sparsity utilization and performance-accuracy trade-off. The 65-nm fabricated chip shows 9.5-TOPS/W system energy efficiency at 4-bit precision, with 6.25 $\times $ actual improvement compared with a state-of-the-art CIM chip. Besides, this work supports high CIM execution accuracy on the ImageNet dataset.
1 February 2018 in Biochemical and Biophysical Research Communications
Voxel-based point cloud networks composed of multiple kinds of sparse convolutions (SCONVs) play an essential role in emerging applications such as autonomous driving and visual navigation. Many researchers have proposed sparse processors for image applications. However, they cannot properly deal with three problems in the point cloud, including low efficiency of random memory access, non-parallel neighbor search and area overhead of supporting hybrid operators, and unbalanced workload among multiple cores. In this work, a 2-D/3-D unified SCONV accelerator is proposed with three key features: a block-wise sparse data storage format supporting out-of-order memory allocation and continuous memory access; a high-throughput and reconfigurable SCONV core providing unified support for multiple kinds of sparse CNNs; an asynchronous and synchronous hybrid scheduler for multiple cores with dynamic on-chip memory router to maximize data reusing and core utilization. This chip is fabricated in 28-nm CMOS technology and achieves 4.68-TOPS/W peak energy efficiency, 2 $\times $ higher than the previous accelerator. It is also the first accelerator to provide unified 2-D/3-D support and end-to-end inference ability for voxel-based point cloud networks.
18 December 2024 in IEEE Journal of Solid-State Circuits
The integration of multi-sensor data like image and point cloud for information complementarity is crucial for 3-D perception scenarios like autonomous driving. Recently, bird’s eye view (BEV)-based sensor fusion is attracting more and more attention but the significant computational overhead constrains their widespread application at the edge. First, there are numerous irregular memory access operations in BEV fusion networks. For example, sparse convolutions (SCONVs) in the point cloud branch and irregular BEV plane mapping result in significant memory addressing and mapping overhead. Furthermore, multi-sensor fusion leads to rapid expansion of model size, making it difficult and expensive for single-chip solutions to meet the demands. Based on the above challenges, this work proposes an image and point cloud fusion processor with two highlights: a content addressable memory (CAM)-based deep fusion core to accelerate a variety of irregular BEV operations and chip-level parallelism design supporting flexible interconnect topology. The proposed chip is fabricated in 28-nm CMOS technology. Compared with existing image or point cloud accelerators, the proposed chip achieves higher frequency, $2\times$ higher area efficiency, and $2.61\times$ higher energy efficiency for sparse point cloud processing. To the best of authors’ knowledge, this work is the first accelerator for BEV-based multi-modal fusion networks.
Wireless image sensor networks acquire and transmit image data through self-organized method. Considering processing ability, energy consumption and cost, this paper implements a practical image sensor node with low-power sleeping mechanism in order to achieve image data's gathering, processing and transmission. To ensure transmission reliability of image data in wireless channel, this paper improves the weakness of retransmission mechanism in traditional Zigbee protocol, and proposes an efficient Zigbee-based retransmission and acknowledgment mechanism. Experimental results show the relationships among the total retransmission time, packet loss rate and the maximum number of retransmission.