Cluster computing has clearly shown the necessity of high performance communication systems. Several high performance network vendors are competing with different approaches: all-in-harware vs. programmable interfaces. Using the programmable Myrinet NICs we show that the performance limits of this approach are far better from the usual expectation when using optimized pipeline software design. Moreover, the ability to propose classical API like MPI with efficient protocols, partially embedded at the NIC level are very promising in terms of performances and communication/computation overlap.