Network-on-Chip (NoC) accelerators with heterogeneous Processing-in-Memory (PIM) cores achieve superior performance than homogeneous ones for neural networks. Dedicated simulators and architecture search frameworks are pivotal for obtaining performance, power, and area (PPA) metrics, as well as guiding the design process. However, existing simulators are primarily designed for homogeneous NoC and lack support for simulating heterogeneous PIM-based NoC architectures. Besides, current search frameworks for heterogeneous NoC architectures only focus on workload allocation and mapping strategies, failing to explore heterogeneous PIM configurations in a larger design space. In this work, we propose HPIM-NoC, a joint simulation and search framework for heterogeneous PIM-based NoC architectures. HPIM-NoC not only supports the simulation of heterogeneous PIM cores, but also provides more accurate latency results by introducing NoC transmission delays and pipelines in co-simulation. HPIM-NoC implements a three-stage heterogeneous search process based on priori knowledge and employs a specific simulated annealing algorithm tailored for heterogeneous architecture search. The search process is accelerated by precomputing core PPA metrics and reducing NoC simulation frequency. In addition, the framework integrates a customized layout algorithm to optimize the placement of heterogeneous NoC, minimizing communication latency and overall area. Experimental results on various neural networks demonstrate that HPIM-NoC can quickly find near-optimal configurations within a limited time. The proposed acceleration method reduces the search time of HPIM-NoC by $2.12 \times$, $2.17 \times$, and $2.96 \times$, respectively. Compared to homogeneous architectures, the Fusions of Metrics (FoMs) of heterogeneous PIM-based NoC architectures found by HPIM-NoC are reduced by $\mathbf{1. 1 8 \%, ~} \mathbf{1 6. 9 4 \%}$, and $\mathbf{3 7. 4 1 \%}$ for ResNet-18 under three settings, respectively.

Hybrid convolutional neural network (CNN) and Transformer networks are emerging in computer vision, combining convolutional, linear, and attention layers to achieve high accuracies with moderate model sizes. Developing the accelerators for hybrid networks is pivotal to simultaneously optimize the static matrix multiplication (MM) in convolutional and linear layers, as well as dynamic MM in attention layers. However, the existing accelerators are primarily designed for either CNNs or Transformers, resulting in increased data movement to support dynamic MM and potential under-utilization of hardware for static MM. To enhance computational performance and energy efficiency for hybrid networks, we propose HyCTor, an accelerator featuring flexible output-stationary (OS) and weight-stationary (WS) dataflows, along with a multicore extension for higher throughput. The parallel array of HyCTor supports interlayer slicing and intralayer splicing to improve the utilization for static MM, and enables seamless switching between OS and WS dataflow to minimize the data movement in dynamic MM. By leveraging structured sparsity in OS dataflow and unstructured sparsity in WS dataflow, the computational efficiency is further boosted for each layer through flexible dataflow selection based on the sparsity ratio. Besides, a novel QuadLoop-mesh topology is proposed to address the complex data dependencies in hybrid networks and minimize data transmission distances in the multicore HyCTor. Experimental results on ResNet-18, ViT-B, and TransIAR-AF show that the proposed single-core HyCTor achieves$1.83\times $,$1.65\times $, and$2.41\times $speedup than state-of-the-art (SOTA) accelerators with 100% utilization rate in most layers, and$3.82\times $–$38.5\times $speedup than RTX4090 GPU. The energy efficiency of HyCTor is improved by$1.81\times $–$8.77\times $compared with SOTA accelerators. Moreover, the 4-core HyCTor achieves speedups of$3.32\times $,$2.58\times $, and$2.91\times $, while the 16-core HyCTor achieves speedups of$7.05\times $,$4.05\times $, and$9.64\times $compared to 1-core HyCTor on three networks.

Large-scale transformer with millions of weights achieves great success in multiple natural language processing (NLP) tasks. To release the memory bottleneck of multi-task model deployment, transfer learning tunes part of weights with shared parameters among tasks. Moreover, computing-in-memory (CIM) emerges as an efficient solution for neural network acceleration. With higher storage density, RRAM-CIM can store the large-scale model without costly weight loading, compared with another mainstream SRAM-CIM. However, the RRAM rewrite for tuned and dynamic weight matrix-vector-multiplication (MVM) in transformers requires high-cost RRAM writing in RRAM-CIM. Current hybrid CIM can compensate for the weakness of RRAM-CIM by adding SRAM-CIM with independent MVM. However, the tuned weights in transfer learning cannot be implemented due to the demand for the cooperative addition of MVM results from both shared and tuned weights. In this paper, a hybrid three-dimension RRAM-CIM and SRAM-CIM architecture (HEIRS) is proposed for multi-task transformer acceleration, with monolithically 3D integration of high-density RRAM-CIM and high-performance SRAM-CIM. The 3D RRAM-CIM with ultra-high density stores the whole model with mitigated off-chip weight loading. The SRAM-CIM is employed for efficiently performing dynamic weight MVM without RRAM rewrite. Moreover, a novel hybrid-CIM paradigm is proposed with an input selective adder tree, to support cooperative addition in transfer learning. Experiments show that, compared with RRAM-CIM and SRAM-CIM, the proposed HEIRS improves the energy efficiency by up to 7.83x and 2.29x on BERT, respectively. Meanwhile, the latency is also reduced by up to 85.5% and the storage density is enhanced by 7.2x, compared to RRAM-CIM.

Modern very-large-scale integrated (VLSI) circuit placement with huge state space is a critical task for achieving layouts with high performance. Recently, reinforcement learning (RL) algorithms have made a promising breakthrough to dramatically save design time than human effort. However, the previous RL-based works either require a large dataset of chip placements for pre-training or produce illegal final placement solutions. In this paper, DeepTH, a three-head policy gradient placer, is proposed to learn from scratch without the need of pre-training, and generate superior chip floorplans. Graph neural network is initially adopted to extract the features from nodes and nets of chips for estimating the policy and value. To efficiently improve the quality of floorplans, a reconstruction head is employed in the RL network to recover the visual representation of the current placement, by enriching the extracted features of placement embedding. Besides, the reconstruction error is used as a bonus during training to encourage exploration while alleviating the sparse reward problem. Furthermore, the expert knowledge of floorplanning preference is embedded into the decision process to narrow down the potential action space. Experiment results on the ISPD 2005 benchmark have shown that our method achieves 19.02% HPWL improvement than the analytic placer DREAMPlace and 19.89% improvement at least than the state-of-the-art RL algorithms.

Hyperspectral imaging (HSI) greatly improves the capacity to identify and monitor ground objects due to the high spectral resolution. As the real-time remote sensing monitoring and warning tasks are getting more attention, new algorithms for low-power on-board classification are required to reduce the transmission time of satellite downlink. In this paper, we propose the Multi-Scale Local Maximum Random Forest (MSLM-RF) to significantly reduce the energy consumption while retaining high classification accuracy. The proposed MSLM-RF uses multi-scale maximum filters for spatial feature extraction and Random Forest for classification after spectral and spatial features fusion. The spatial features are efficiently extracted with low computational complexity by regarding the maximum light intensity values in different ranges of pixels as anchor points. MSLM-RF only consists of integer comparisons and a few additions, thereby eliminating the energy-hungry operations such as multiplication and exponentiation. According to experimental results on the HSI benchmark datasets, MSLM-RF delivers a better trade-off in accuracy and computational complexity than the state-of-the-art classification algorithms. Besides, MSLM-RF gets higher average classification accuracy and lower energy consumption than the previous on-board algorithms. The obtained results show the suitability of the proposed algorithm to accomplish practical real-time classification tasks on-board with low energy consumption.