Deep Neural Networks (DNNs) have demonstrated their utility across a wide range of input data types, usable across diverse computing substrates, from edge devices to datacenters. This broad utility has resulted in myriad hardware accelerator architectures. However, DNNs exhibit significant heterogeneity in their computational characteristics, e.g., feature and kernel dimensions, and dramatic variances in computational intensity, even between adjacent layers in one DNN. Consequently, accelerators with static hardware parameters run sub-optimally and leave energy-efficiency margins unclaimed. We propose DyHard-DNNs, where accelerator microarchitectural parameters are dynamically reconfigured during DNN execution to significantly improve metrics of interest. We demonstrate the effectiveness of this approach on a configurable SIMD 2D systolic array and show a 15--65% performance improvement (at iso-power) and 25--90% energy improvement (at iso-latency) over the best static configuration in six mainstream DNN workloads.

Massively parallel computation in GPUs significantly boosts performance of compute-intensive applications but creates power and thermal issues that limit further performance scaling. This paper demonstrates significant GPGPU power savings by relaxing application accuracy requirements and enabling the use of low power imprecise hardware (IHW). A synthesized set of novel imprecise floating point arithmetic units is presented. GPGPU-Sim and GPUWattch are used to estimate impacts of IHW units on output quality and system-level power consumption, providing a quality-power tradeoff model for application-specific optimization. Experimental results for a 45 nm process show up to 32% power savings with negligible impacts on output quality.

Recent imprecise hardware (IHW) design methodologies present opportunities for achieving gains in nonfunctional efficiency design metrics by allowing errors in computation within error tolerant application domains. This work presents a novel imprecise Error Tolerant Balancing Adder (ETBA) design - an augmentation of the ETAIIM IHW adder that reduces errors by introducing a balance block that detects and corrects carry chain inconsistencies in the ETAIIM but operates off the critical path. Furthermore, this work identifies a common class of killer inputs with high error rates for IHW adders, and demonstrates the ETBA's resilience to these errors. A JPEG decompression case study reveals a 24% reduction in ETBA addition energy-delay product compared to a Kogge-Stone adder with only a 0.2% reduction in SSIM image quality.

Subthreshold digital circuits minimize energy per operation and are thus ideal for ultralow-power (ULP) applications with low performance requirements. However, a large range of ULP applications continue to face performance constraints at certain times that exceed the capabilities of subthreshold operation. In this paper, we give two different examples to show that designing flexibility into ULP systems across the architecture and circuit levels can meet both the ULP requirements and the performance demands. Specifically, we first present a method that expands on ultradynamic voltage scaling (UDVS) to combine multiple supply voltages with component level power switches to provide more efficient operation at any energy-delay point and low overhead switching between points. This system supports operation across the space from maximum performance, when necessary, to minimum energy, when possible. It thus combines the benefits of single-VDD, multi-VDD, and dynamic voltage scaling (DVS) while improving on them all. Second, we propose that reconfigurable subthreshold circuits can increase applicability for ULP embedded systems. Since ULP devices conventionally require custom circuit design but the manufacturing volume for many ULP applications is low, a subthreshold field programmable gate array (FPGA) offers a cost-effective custom solution with hardware flexibility that makes it applicable across a wide range of applications. We describe the design of a subthreshold FPGA to support ULP operation and identify key challenges to this effort.

The energy efficiency of a CMOS architecture processing dynamic workloads directly affects its ability to provide long battery lifetimes while maintaining required application performance. Existing scalable architecture design approaches are often limited in scope, focusing either only on circuit-level optimizations or architectural adaptations individually. In this paper, we propose a circuit/architecture co-design methodology called Panoptic Dynamic Voltage Scaling (PDVS) that makes more efficient use of common circuit structures and algorithm-level processing rate control. PDVS expands upon prior work by using multiple component-level PMOS header switches to enable fine-grained rate control, allowing efficient dithering among statically scheduled algorithms with sub-block energy savings. This way, PDVS is able to achieve a wide variety of processing rates to match incoming workload as closely as possible, while each iteration takes less energy to process than on architectures with coarser levels of rate control. Measurements taken from a fabricated 90 nm test chip characterize both savings and overheads and are used to inform PDVS synthesis decisions. Results show that PDVS consumes up to 34% and 44% less energy than Multi-VDD and Single-VDD systems, respectively.

Dynamic voltage scaling (DVS) provides power savings for systems with varying performance requirements. One low overhead implementation of DVS uses PMOS power switches to connect DVS blocks to one of the available VDDsupplies. While power switches have been analyzed extensively for leakage power gating, proper design of power switches for DVS is less well understood. This paper characterizes power switches for DVS in terms of VDD-switching delay and VDD-switching energy. We show the impact of these switching overheads on a novel fine-grained DVS architecture and present an RC model that allows fast estimation of the overhead. Measurements of a DVS multiplier and adder on a 90 nm CMOS test chip validate the model. Our model and measurements confirm that power switched DVS can provide sufficiently low overhead to give energy savings with only one clock cycle spent at a lower voltage, making this approach a flexible and enticing option for embedded portable systems.