In recent times, detection of mental stress using physiological signals has received widespread attention from technology researchers. Although many interesting research works have been reported in this area, evidence of hardware implementation is occasional. The main challenge in stress detection research is use of the optimum number of physiological signals and real-time detection with low-complexity algorithms. In this work, a real-time stress detection technique is presented which utilizes only a photoplethysmogram (PPG) signal to achieve improved accuracy over multi-signal-based mental stress detection techniques. A short 5 s segment of a PPG signal was used for feature extraction with an autoencoder (AE), and features were minimized using recursive feature elimination (RFE) integrated with a multi-class support vector machine (SVM) classifier. The proposed AE-RFE-SVM-based mental stress detection technique was tested with the WeSAD dataset to detect four levels of mental state, namely baseline, amusement, meditation and stress, and achieved an overall accuracy, F1 score and sensitivity of 99%, 0.99 and 98%, respectively, for 5 s PPG data. The technique provided improved performance over discrete wavelet transformation-based feature extraction followed by classification with either of the five types of classifiers, namely SVM, random forest, k-nearest neighbor, linear regression and decision tree. The technique was translated into quad-core-based standalone hardware (1.2 GHz, 1 GB RAM). The resultant hardware prototype achieved low latency (∼ 0.4 s) and had a low memory requirement (∼ 1.7 MB). The present technique can be extended to develop a remote healthcare system using wearable sensors.