<p>Driver drowsiness detection is a latency-critical, computation-intensive challenge in intelligent transportation systems, where the ability to capture subtle transitions between alert and drowsy states must occur within sub-second time windows for real-time safety. Meeting this requirement demands high-performance computing (HPC) primitives—multi-threaded frame acquisition, parallel batched inference, and fine-grained relational learning—that go far beyond the capacity of conventional sequential pipelines. This paper presents a novel deep metric learning framework based on a septuplet convolutional neural network, designed to identify and rank nuanced stages of driver drowsiness using a one-shot learning approach. The proposed model processes seven images simultaneously and employs a septuplet loss function to learn highly discriminative features by maximizing inter-class distances while minimizing intra-class variation. Because each training step propagates seven parallel forward passes through a shared encoder, the workload maps naturally onto data-parallel SIMD/GPU pipelines, and we exploit Python multi-threading at inference to decouple frame acquisition from neural computation, sustaining 156 frames per second on commodity hardware. The system is evaluated using the NTHU Driver Drowsiness Detection Dataset under a leave-one-subject-out (LOSO) cross-validation protocol, and its generalization is further verified on the UTA-RLDD dataset. Unlike traditional binary classification, our model classifies input into five distinct drowsiness classes: pure drowsy, pure non-drowsy, similar drowsy, partial drowsy, and partial non-drowsy, allowing for a more granular assessment of driver state. By integrating more ranking constraints without overloading the computational cost, the septuplet network achieves an optimal balance between accuracy and efficiency, making it highly scalable and suitable for real-time, parallel, and embedded deployment in safety-critical transportation systems.</p>