There is much interest in embedding data analytics into sensor-rich platforms such as wearables, biomedical devices, autonomous vehicles, robots, and Internet-of-Things to provide these with decision-making capabilities. Such platforms often need to implement machine learning (ML) algorithms under stringent energy constraints with battery-powered electronics. Especially, energy consumption in memory subsystems dominates such a system's energy efficiency. In addition, the memory access latency is a major bottleneck for overall system throughput. To address these issues in memory-intensive inference applications, this dissertation proposes deep in-memory accelerator (DIMA), which deeply embeds computation into the memory array, employing two key principles: (1) accessing and processing multiple rows of memory array at a time, and (2) embedding pitch-matched low-swing analog processing at the periphery of bitcell array. The signal-to-noise ratio (SNR) is budgeted by employing low-swing operations in both memory read and processing to exploit the application level's error immunity for aggressive energy efficiency.This dissertation first describes the system rationale underlying the DIMA's processing stages by identifying the common functional flow across a diverse set of inference algorithms. Based on the analysis, this dissertation presents a multi-functional DIMA to support four algorithms: support vector machine (SVM), template matching (TM), k-nearest neighbor (k-NN), and matched filter. The circuit and architectural level design techniques and guidelines are provided to address the challenges in achieving multi-functionality. A prototype integrated circuit (IC) of a multi-functional DIMA was fabricated with a 16 KB SRAM array in a 65 nm CMOS process. Measurement results show up to 5.6X and 5.8X energy and delay reductions leading to 31X energy delay product (EDP) reduction with negligible (<1%) accuracy degradation as compared to the conventional 8-b fixed-point digital implementation optimally designed for each algorithm.Then, DIMA also has been applied to more complex algorithms: (1) convolutional neural network (CNN), (2) sparse distributed memory (SDM), and (3) random forest (RF). System-level simulations of CNN using circuit behavioral models in a 45 nm SOI CMOS demonstrate that high probability (>0.99) of handwritten digit recognition can be achieved using the MNIST database, along with a 24.5X reduced EDP, a 5.0X reduced energy, and a 4.9X higher throughput as compared to the conventional system. The DIMA-based SDM architecture also achieves up to 25X and 12X delay and energy reductions, respectively, over conventional SDM with negligible accuracy degradation (within 0.4%) for 16X16 binary-pixel image classification. A DIMA-based RF was realized as a prototype IC with a 16 KB SRAM array in a 65 nm process. To the best of our knowledge, this is the first IC realization of an RF algorithm. The measurement results show that the prototype achieves a 6.8X lower EDP compared to a conventional design at the same accuracy (94%) for an eight-class traffic sign recognition problem.The multi-functional DIMA and extension to other algorithms naturally motivated us to consider a programmable DIMA instruction set architecture (ISA), namely MATI. This dissertation explores a synergistic combination of the instruction set, architecture and circuit design to achieve the programmability without losing DIMA's energy and throughput benefits. Employing silicon-validated energy, delay and behavioral models of deep in-memory components, we demonstrate that MATI is able to realize nine ML benchmarks while incurring negligible overhead in energy (< 0.1%), and area (4.5%), and in throughput, over a fixed four-function DIMA. In this process, MATI is able to simultaneously achieve enhancements in both energy (2.5X to 5.5X) and throughput (1.4X to 3.4X) for an overall EDP improvement of up to 12.6X over fixed-function digital architectures.
PDF
download
878987797
028-85220240
