Intramuscular electromyography (iEMG) decomposition identifies motor neuron (MN) discharge timings from interference iEMG recordings. When this is performed in real-time, the extracted neural information can be used for establishing human-machine interfaces. We propose a multi-channel real-time decomposition algorithm based on a Hidden Markov Model of EMG and a Bayesian filter to estimate the spike trains of motor units (MUs) and their action potentials (MUAPs). The multi-channel framework of Bayesian modelling and filtering was implemented into parallel computation using multiple GPU clusters, which ensures computational speed compatible with real-time decomposition. A decomposed-checked channel strategy is then proposed for arranging channels into groups to be processed in related GPU clusters. The algorithm was validated on six 16-channel simulated signals, three 32-channel experimental signals acquired from the human tibialis anterior muscle, and two 16-channel experimental signals acquired from the abductor digiti minimi muscle with thin-film implanted electrodes. All signals were decomposed in real time with an average decomposition accuracy > 90%. In conclusion, the proposed multi-channel iEMG decomposition algorithm can be applied to implanted multi-channel electrode arrays to establish human-machine interfaces with high-information transfer.