Panel: How to Compute with a Billion Transitors on a Chip

Although these panels are traditionally more entertainment-oriented than information-oriented, this one was actually more serious...

Tom Knight (MIT)

Transistors are not the problem. Wires are the problem. To reduce the wires, let's explore the possiblity of attaching more sophiscated functional units at the end of the wires.

Hank Levy (University of Washington)

Threads are the solution. Instead of working hard to schedule code from a "bad" thread (one that has reached a "dead end"), let's have lots of threads, and we'll always find code to schedule from the "good" threads.

Henry Massalin (MicroUnity)

We know many tricks, each of which might give us some moderate improvement. Examples are threads, vectors, and multi-issue. Since we have so many transistors to play with, let's do all! The combined benefit gotta be bigger than that of any individual approach.

Kunle Olukotun (Stanford University)

Instead of devoting transistors to huge caches or building extremely complicated uniprocessors, let's put multiple simple processors on a single chip (see technical session #1).

Bob Parker (ARPA)

The buzz word is "temporal reuse of silicon". They want to be able to dynamically reconfigure the silicon to do different things.

Guri Sohi (University of Wisconsin)


Previous                    Home                     Next


Randy Wang ( rywang.public@gmail.com )

This page created with Netscape Navigator Gold (last updated on October 10, 1996)