Samsung is one of the biggest consumer device companies, manufacturing devices from digital home appliances to mobile smartphones. Providing the best user experience is our utmost R&D goal, which is technically challenging because of the varieties of devices spanning from low-cost appliances to high-end phones. To achieve this goal, we strive for end-to-end system optimization by co-designing workload analysis, system SW development, and SoC HW design.
Our research approach comprises three pillars, Customized IP (Intellectual Property), SoC Architecture, and AI Framework. In particular, we focus on on-device AI. We develop specialized HW accelerators, integrate them into SoCs seamlessly, and provide compiler and runtime frameworks to developers. Through our research, we will provide Samsung devices AI functionalities as comparable to cloud-based AI services, while providing lower latency and higher privacy.
Our vision is to provide the best user experience across Samsung devices by optimizing end-to-end system optimization. By co-designing algorithms, software, and hardware, we optimize SoC architecture to achieve the best power and performance.
We develop customized IPs with our proprietary algorithms for product differentiation.
For example, we developed and productized an NPU (Neural Processing Unit), demonstrating excellent power, performance and area among many other commercial NPU IPs. Further, SW Toolchain is crucial and we also developed simulator, compiler, runtime, and device driver for our NPUs. For future NPU development, we research deeply on AI model compression algorithms. The research will enable us to identify latest AI algorithms and specialize our NPU architecture for them.
We research custom SoC architecture to provide seamless user experience across Samsung devices.
We strive for an end-to-end optimization from user applications, through Tizen & Android OS, to the HW platform. By analyzing user application workloads, we understand their computing and memory characteristics. Based on the acquired knowledge, we first identify the power and performance bottlenecks of our systems, then we optimize our system software and architect our hardware platform accordingly.
We develop and publish our on-device AI frameworks such as AI pipelining, on-device learning, and AI micro-runtime.
On-device AI frameworks are becoming more powerful and complex, with an increasing variety of AI applications and services. We need to manage data, models, and application deployments, similar to what ML Operations (MLOps) targets, but in different environments and requirements of embedded devices. We develop “Device MLOps” that act as MLOps on devices, provide application interfaces, and collaborate with “Traditional” MLOps. We develop unified AI frameworks that can be commonly applied to various devices of Samsung as well as other companies.