Is it possible to combine your approach with NATTEN? It seems that both approaches are optimizing from different directions and can be combined with significant throughput and small performance improvements?