For all three categories of machines, there is significant freedom in designing the processors. The results of design choices will have little impact on the final costs of the machines, but could have a very large impact on the performance. Given that it is possible to design a PetaFLOPS machine, the implementation depends on successes in other areas. Device technology has to advance in the years after 2007. Programming technology has to develop paradigms and tools that scale to the development of programs for 10,000 processor machines. Applications programs have to based on models that lend themselves well to partitioning or streaming, and that can be cast in forms compatible with latency hiding and latency removal.
These challenges should result in feedback to the architecture community that helps guide the details of what can and should be put into future processors. The possible choices are virtually limitless. The relative costs of one method versus another may be very small. What is not known is which choices are good and which choices are not. To evaluate the choices, the architects need guidance from the device technologists regarding the characteristics of devices that will become available, and from the applications and systems programmers regarding what features make programs more efficient on machines in the PetaFLOPS range. Working together, the various teams should be able to achieve significant progress in the years to come.