We proposed a low energy-delay-product (EDP) computing-in-memory (CIM)-based human pose estimation (HPE) accelerator in mobile VR devices by computing depth-wise separable convolution of a lightweight HPE network HW-effectively in CIM. The processor achieved high throughput and energy-efficiency compared to the previous SoTA CIM. Therefore, it operates HPE with a low EDP of 27.6 uJ ⋅ s in mobile VR devices.