Anchor-based Gaussian Splatting has demonstrated great potential to improve the NVS rendering speed, quality and efficiency. However, achieving resource-efficient training and rendering on edge devices is still challenging. In this work, we first perform extensive profiling and identify the anchor GS bottlenecks on redundant anchor-wise operations and hardware under-utilization. To address these problems, we present CFoM-GS, a unified training-rendering accelerator for anchor-based Gaussian Splatting. A structural pruning scheme is developed to trim occluded or unimportant anchors and Gaussians in early stage of rendering. A set of chunk-wise rendering cores are designed to make use of the multi-chunk workload scheduling to render footprint-matched Gaussians simultaneously. Our CFoM-GS is implemented in commercial 22nm technology and is evaluated by multiple real-world 3D scenes. Compared to edge Jetson NX GPU, CFoM-GS achieves 22× throughput improvement and three orders of magnitude energy efficiency gain.

Embodied artificial intelligence (Embodied AI) bridges digital intelligence and physical interaction, promising transformative applications in robotics and AGI. While vision-language-action (VLA) models have revolutionized embodied agents with unprecedented perception and reasoning capabilities, their real-world deployment is bottlenecked by high inference latency and the absence of efficient edge acceleration. In this work, we present a heterogeneous accelerator to address the varying computing pattern of modern VLAs. We begin with an in-depth profiling of state-of-the-art VLA models, identifying computation bottleneck in VLM and latency bottleneck in DiT denoising process. Based on the observations, we utilize a heterogeneous architecture to accommodate both the VLM and the DiT computations, together with an iteration-aware low-rank approximation scheme to accelerate the linear layers in DiT. Based on extensive evaluations, our solution delivers up to 7.2× end-to-end speedup and 201× energy efficiency boost compared to mobile GPU RTX-3060 and outperforms SOTA accelerators by 14.5× and 6.2× higher area and energy efficiency.

Subgraph Graph Neural Networks (GNNs) are emerging as a promising approach to enhance GNN expressiveness, but their more complex graph structures with numerous independent and irregular subgraphs pose significant hardware deployment challenges. In this work, we propose 3D-SubG, a 3D stacked hybrid processing-near/in-memory accelerator for subgraph GNNs. With hybrid bonding packaging technology, a logic die is 3D stacked with a DRAM die for highly parallel memory accesses. The logic die employs digital SRAM-based processing-in-memory (PIM) macros to boost computation density and minimize data transfer. We further propose a bit-level non-zero gathering method to exploit graph sparsity for PIM, a workloadbalanced mapping strategy for subgraph allocation onto different logic-to-DRAM blocks, and a distributed global pooling approach to reduce inter-block data movements. Experimental results show that 3D-SubG achieves average improvements of $146.11 \times$ in performance, $934.18 \times$ in area efficiency, and $1171.80 \times$ in energy efficiency compared to RTX 3090Ti.