I do not agree. When you benchmark you usually measure the differences between times needed to complete. This is because it is highly non-obvious that if you increase workload twice, the time increases also twice. Perhaps the algorithm is not linear. Perhaps if you have more data, you suddenly need to swap memory. Perhaps something (like disk access in parallel) means that actually it takes less than 2x time. This means that a concept of speed per unit of work is undefined. So the only reasonable interpretation of "70% faster" means "spends 30% time of original".