The author trained a neural network on dual AMD MI50 32GB GPUs using Vulkan as a compute API, despite AMD ending official ROCm support for the gfx906 architecture after ROCm 5.7. This proof‑of‑concept shows that serious AI workloads can run on legacy MI50 hardware without ROCm by leveraging Vulkan‑based frameworks. The approach counters the prevailing view that newer GPUs are required for AMD‑based deep learning.