D Gaussian Splatting (3DGS) has emerged as a promising real-time photorealistic radiance field rendering technique. Existing GPU and hardware accelerators face limitations due to insufficient parallelism in sequential rendering pipeline stages and the memory overhead associated with interim results. This paper presents GSAcc, a hardware accelerator co-designed with dataflow to render compressed 3DGS models on edge platforms efficiently. GSAcc enhances 3DGS rendering performance through several key innovations. First, it introduces Gaussian depth speculation, parallelizing preprocessing and sorting tasks. Second, GSAcc adopts a Gaussian-centric dataflow that interleaves preprocessing and rasterization, allowing all rendering steps to execute concurrently without storing intermediate results. Finally, it employs dedicated hardware acceleration to address sorting and rasterization bottlenecks within the optimized dataflow. We implemented and synthesized GSAcc using Intel16 PDK and evaluated its performance on real-world 3DGS scenes. Compared with desktop GPUs, GSAcc achieves up to $1.66 \times 10^{4} \mathrm{x}$ Power-Performance-Area (PPA) improvement as well as 48.7 x energy savings. Additionally, GSAcc outperforms the state-of-the-art hardware accelerator GSCore with up to 2.3x PPA improvement and 2.9x energy savings.

The computational demands for recent large transformer- based language models and Neural Radiance Fields (NeRF) have rapidly increased, impacting applications like conversational AI and Mixed Reality (MR). Current accelerator architectures struggle to cope with the vast computational requirements, creating a gap with slowly growing hardware resources. This paper proposes repurposing memory components as high-density computational units, leveraging recent advancements in Back-End-Of-Line (BEOL) transistors and monolithic 3D integration techniques. An ultra-high density monolithic 3D eDRAM is presented as a reconfigurable matrix multiplication unit, co-designed with analog computation circuits, achieving energy efficiency up to 2.41 TOPS/W, performance up to 1.71 TOPS on bfloat16, and compute intensity up to 118 GOPS/mm2. A comprehensive multi-cube(core) architecture is also devised and optimized with bit stationary tensorcore dataflow. We evaluate the proposed architecture on state-of-the-art machine learning models: NeRF and LLaMa-7B, improving the computation density by up to 6.59x and 1.12x compared with GPU and state-of-the-art vector processor designs, respectively.