Aphelios: A Selective Lock-step Neural Processing Unit Design

Engineering notes

Traditional ways of ensuring reliability, such as resource duplication and lock-step design, will introduce significantly high area and energy overhead. Using the well-adapted Resiliency Accuracy (RA) metric, Aphelios achieves only 0.4% accuracy loss compared to a full system duplication lock-step NPU design.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Neural Processing Units (NPU) that are widely used in autonomous machines are not reliable under soft errors. A random single-bit flip happening in an NPU can easily result in a malfunction of the perception or localization module in a self-driving vehicle. Traditional ways of ensuring reliability, such as resource duplication and lock-step design, will introduce significantly high area and energy overhead. In this work, we propose a selective lock-step NPU design for neural network reliability. By duplicating only a small portion of the circuits and comparing only the important neurons, we can still guarantee high reliability with a relatively low overhead. Using the well-adapted Resiliency Accuracy (RA) metric, Aphelios achieves only 0.4% accuracy loss compared to a full system duplication lock-step NPU design. At the same time, Aphelios achieves 70.1% less overhead on area and 68.1% less energy overhead.

Links and sources

• Official / arXiv page

Comments