unified memory architectures are getting more interesting for inference workloads.