It won't be linear for memory-bound applications. The v100 was able to make it close to linear with large enough transfer sizes, but it has 50% more memory bandwidth than these.