A multi-year Caltech project developed 

 high performance interconnect and fine- 

 grained parallel systems including 

 boards consisting of 64 single nodes, 

 scalable to thousands of nodes. 

 Several commercial systems have 

 adopted architectures based on this 

 research, and are beginning to demon- 

 strate how low cost modules can be 

 configured to meet a broad range of 

 applications. 



Joint projects with other agencies are established to 

 accelerate technology development and transition. 

 ARPA joint projects with NSF include foundations for 

 scalable systems, visualization, Grand Challenges, giga- 

 bit networks, and accelerating the maturation of systems 

 software at NSF Supercomputer Centers. Joint projects 

 with NASA include an Internet software exchange, sys- 

 tem software maturation, and ground stations for the 

 ACTS gigabit satellite system. ARPA, NSF, and NASA 

 also have a joint program in digital libraries. Joint pro- 

 jects with DOE include scalable software libraries and 

 networking applications. ARPA is working with NIST to 

 develop performance measurement technologies and 

 techniques, privacy and trusted systems technologies, 

 and the computer emergency response team system for 

 the Internet. A joint project with NSA is developing giga- 

 bit network security technology and other secure and 

 trusted systems technologies. In addition, a variety of 

 early evaluation and experimental use projects involve 

 different kinds of scalable parallel computing systems. 



The ARPA program focuses on the advanced technolo- 

 gy aspects of all five components of the HPCC Program 

 as follows: 



HPCS 



ARPA projects stimulate the development of scalable 

 computing technologies that are capable of being con- 

 figured as networks of workstations and large scale par- 

 allel computing systems capable of sustaining trillions of 

 operations per second. Systems can be configured over 

 a wide performance range. The systems will be bal- 

 anced to provide the processor-to-memory, scalable 

 interconnection, and input/output bandwidth needed to 

 sustain high internal and external system performance. 

 The modular design of the system units of replication 

 will enable them to cover the full range from worksta- 

 tions to the largest scale distributed and parallel sys- 

 tems. Scalable systems with vector accelerators may be 

 configured as parallel vector systems. Other kinds of 

 accelerators such as field programmable logic arrays 

 may be added for specialized applications. The largest 

 scale parallel systems with hundreds to thousands of 

 processors or more, are sometimes referred to as mas- 

 sively parallel systems. The input/output interfaces of 

 these systems may be used to configure heterogeneous 

 systems with high performance networks. 



60 



