The director of architecture for MIPS Technologies, and the person directly responsible for the company's recently announced signal-processing extensions, takes the pragmatic view. "Sure, there's architectural elegance that you strive for," said Radhika Thekkath. "But much of the job is interacting with implementers — our designers inside MIPS and the SoC [system-on-chip] designers at our silicon partners.

"A lot of architecture is recognizing the problems that system designers face and finding ways to help them," Thekkath went on. "There is a lot you can do. The base architecture — the pure part — is shielded from the rest of the system by an intermediate layer, the microarchitecture. So you try to keep the base architecture clean and adapt it to the system through the microarchitecture."

Increasingly, those interactions must include not just hardware implementers but compiler and applications designers as well. "Software is one of the keys to the system," Thekkath said. "Nice feature, but can the software developers actually use it? The SIMD instructions we recently announced are a good illustration. Without an adequate vectorizing compiler, only people who can hand-code SIMD loops could use the extensions. So you have to include the compiler and applications developers in your thinking."

Thekkath received a PhD in computer science, not electronics engineering, from the University of Washington. "The university has a very systems-oriented program," she said. "Computer science there was not a theoretical discipline at all but had operating systems and compiler technology tracks as well. My thesis, as it happened, was related to multithreading — that was in the very early days of the work on the topic."

After graduation, Thekkath spent two years working for RISC pioneer John Hennessy at Stanford. Then, after a stint at Equator Technology, she joined MIPS. Now she is the keeper of the MIPS architecture. But she is also an explorer, a listener, a purveyor of "what ifs." To Thekkath, building the modern architecture is not constructing an edifice in the clouds but finding ways to make the fundamentals of the MIPS architecture more valuable to real designers — hardware and software — in real systems.