AnaCoNGA: Analytical HW-CNN Co-Design Using Nested Genetic Algorithms

MARLIN: A Co-Design Methodology for Approximate ReconfigurabLe Inference of Neural Networks at the Edge

“…NN accelerators are typically based on a hardware architecture that comprises a memory hierarchy and an array of interconnected processing engines (PEs) [4], [7], [8], [9], [10], similar to the one depicted in Figure 1. The memory hierarchy usually comprises the system's main memory (offchip), global buffers (on-chip), and the registers within the processing engines.…”

“…Figure 10 depicts the utilization of approximation levels for each NN layer. It is possible to notice the presence of peaks around levels with an index equal to or smaller than 2 j − 1, j ∈ [0, 1,2,3,4,5,6,7,8], corresponding to the Pareto optimal points of Figure 8. In Figure 10, it is shown how the majority of approximation levels are used in the Pareto front, justifying the choice to maintain power and MRED-dominated configurations as the most efficient levels might not be optimal ones to achieve a high task accuracy.…”

“…Digital Object Identifier 10.1109/TCSI.2024.3365952 challenges in deploying them on energy-constrained devices, such as microcontrollers (MCUs) [1], [2]. Over the past decade, dedicated hardware accelerators [3], [4], [5], [6], [7], [8], [9] have been developed to optimize energy efficiency and throughput during inference by reducing the data movement and the cost of arithmetic operations. The latter usually accounts for around a fourth of the total inference energy, as the convolutional and fully connected layers of NNs involve millions of multiplications and additions [10].…”

“…It lowers the amount of memory traffic and the computational cost, allowing the deployment of NNs on low-power devices with no support for floatingpoint arithmetic. Layer-wise quantization leverages the different degrees of robustness and tolerance to error introduction of NN layers.Therefore, a runtime reconfigurable multiplier supporting operands with different bitwidths is necessary to leverage mixed-precision and layer-wise quantization [3], [4], [11], [12]. Similarly to quantization, approximation is another co-design technique mainly aimed at reducing the inference energy [14].…”