Core Viewpoint - The article introduces the embedded GPU (e-GPU), a configurable RISC-V GPU platform designed specifically for ultra-low-power edge devices (TinyAI), addressing the challenges of power consumption and area constraints in traditional GPU implementations [1][6]. Group 1: Introduction and Background - The increasing demand for real-time computing driven by machine learning is propelling the rapid development of edge computing, which enhances privacy and energy efficiency by processing data locally rather than relying on cloud servers [4]. - Specialized hardware architectures are required to meet the performance, real-time response, and power consumption limitations of these workloads, with heterogeneous architectures integrating CPUs and domain-specific accelerators being an effective solution [4][5]. - Traditional GPUs have not been thoroughly studied for their trade-offs in ultra-low-power edge devices, which typically operate under strict power constraints in the tens of milliwatts range [5][6]. Group 2: e-GPU Architecture and Features - The e-GPU architecture is designed to minimize area and power consumption while being adaptable to TinyAI applications, featuring a configurable design that allows for optimization of area and power [24][25]. - The memory hierarchy employs a unified architecture that maps the host's main memory and e-GPU global memory to the same physical memory, enhancing programmability and reducing data transfer complexity [26][27]. - A dedicated controller manages e-GPU operations, integrating power management features to monitor and control the power state of computation units [29]. Group 3: Performance Evaluation - The e-GPU configurations were tested using two benchmark tests: General Matrix Multiplication (GeMM) and TinyBio, demonstrating significant performance improvements and energy savings [48][49]. - The e-GPU system achieved speedups of up to 15.1 times and energy reductions of up to 3.1 times compared to baseline systems, while maintaining a power budget of 28 mW [2][58]. - The area of the e-GPU system ranged from 0.24 mm² to 0.38 mm², proving its feasibility for deployment in TinyAI applications, which typically have strict area constraints [50]. Group 4: Industry Context - Commercial edge GPU solutions, such as Qualcomm's Adreno and ARM's Mali GPUs, are not specifically designed for TinyAI applications, often exceeding the power requirements needed for these applications [11]. - Academic GPU research focuses on developing programmable and configurable architectures suitable for various computing domains, with the e-GPU proposed as a suitable solution for TinyAI workloads [12][13]. - The e-GPU platform is positioned as an open-source, configurable RISC-V GPU platform that addresses the programming limitations and energy efficiency needs of the TinyAI domain [12][13].