This is indeed a good moel. From the article's context, dGPUs have more of both bandwidth and flops, the compute-intensity balance isn't necessarily the decidin

This is indeed a good moel. From the article's context, dGPUs have more of both bandwidth and flops, the compute-intensity balance isn't necessarily the deciding factor on whether there's a speedup on GPU. In this article the deciding factor seems to be the startup cost because the application has placed the data on the CPU memory side, and is considering shipping out to GPU memory just for this computation.