10 August 2023 in IEEE Transactions on Circuits and Systems I: Regular Papers
Compute-in-memory (CIM) is a promising approach to solving the memory-wall problem existing in traditional computing architectures. In this paper, we introduce SSM-CIM, a charge-domain, static random-access memory (SRAM)-based CIM macro designed for area-energy-efficient convolutional neural network (CNN) inference. SSM-CIM utilizes an original sign-magnitude data encoding method for both inputs and weights. By codesigning four adjacent SRAM computing cells and employing a 3-bit digital-to-analog converter (DAC), SSM-CIM performs accurate 4-bit multiply-and-accumulate (MAC) computation in a single step, eliminating the peripheral digital shift-and-add circuits. To digitize the MAC computing results, a dedicated multi-reference assisted SAR ADC is designed by reusing the reference voltages from the DAC, which offers significant power and area savings. In addition, analog computing errors and quantization errors are analyzed to ensure the multi-bit computing accuracy of SSM-CIM. SSM-CIM is implemented and evaluated using 28-nm global foundry process. The post-layout simulation results validate the excellent computing linearity and accuracy of SSM-CIM. Benefitting from the compact layout design and fully parallel computing flow, the $144\times 256$ macro achieves a peak throughput of 2.3 TOPS, an area efficiency of 10.2 TOPS/mm2, and an energy efficiency of 205.4 TOPS/W with 4-bit weights and 4-bit inputs.
12 August 2024 in Journal of Lightwave Technology
Optical switching based on arrayed waveguide grating routers (AWGRs) and fast tunable sources is a future-proof solution to overcome the bottleneck of limited bandwidth and high latency of electrical switches-based data center networks (DCNs). To support fine-grained optical packet switching and maintain high network throughputs, tunable sources capable of nanosecond-scale hitless wavelength switching are critical requirements for such fast optically switching networks. In this work, we experimentally demonstrate a fast-tunable source based on a high-density multi-wavelength DFB laser array monolithically integrated with a three-stage Y-branch combiner and an SOA. The DFB laser array has an average 100-GHz channel spacing under the same bias current. The output power of all lasers is above 10 mW, with a uniformity of less than 1 dB. Benefitting from the high-speed dynamic property of the DFB laser and high-speed burst drive pulses controlled by a field-programmable gate array (FPGA), less than 3 ns switching time is achieved by the on-off switching of the DFB laser on arrays, which is independent of tuning range. Furthermore, a 2 × 2 optical switching prototype based on this demonstrated chip and AWGRs achieves error-free transmission atop 25 Gbps NRZ per link with 12.8 ns end-to-end reconfiguration, which provides a path toward a high-bandwidth, low-latency, and energy-efficient wavelength-switched network for data centers in the post-Moore's Law era.