What I'd prefer over benchmarks is the answer to a simple question:

What useful thing can it demonstrably do that its predecessors couldn't?