I believe there are two main things holding it back. One is an impoverished execution model, which makes certain tasks difficult or impossible to do efficiently

I believe there are two main things holding it back. One is an impoverished execution model, which makes certain tasks difficult or impossible to do efficiently; GPUs … struggle when the workload is dynamic This sacrifice is a purposeful cornerstone of what allows GPUs to be so high throughput in the first place.