One of the most important challenges facing computing clusters in the foreseeable future is providing fault tolerant, high availability cluster hardware for non-stop applications. This capability is in addition to high throughput and low latency. This paper presents the Advanced MultiProcessor Network (AmpNet), a gigabit speed cluster interconnect that was designed with these issues in mind. The AmpNet Network Interface Card (NIC) uses network-shared memory as network cache to provide a fault-tolerant, self-healing network with no data loss. Higher-level network centric services use network-shared memory to ensure high availability and continuity of service in applications. In addition, the programmable NIC, with low-latency messaging protocols and field upgradeable soft logic, provides a foundation for researchers who would like to develop additional cluster services and protocols for network centric computing. This paper describes the fault tolerant design and implementation of the AmpNet hardware architecture.