25%-80%.

Certain (mostly AI) workloads. Depends on how hot you're comfortable running your card. The real-world improvement is considerably less due to other things being bottlenecks.