A Glimpse of VLIW Research at IBM

by Bradley C. Kuszmaul

I'll present my understanding of the superscalar processor architecture being developed by the advanced VLSI architecture research group, lead by Kemal Ebcioglu, at IBM T.J. Watson Research Center in Yorktown. I visited IBM in March, and talked to Kemal and his team, who explained their architecture to me. They claim to get 5 instructions per clock on a wide range of C code. The chief innovation, as I see it, is the use of static multiway predicated instructions. Each instruction consists of up to 8 instructions that are each conditionally executed based on several condition registers. To explain the architecture, I'll show how a simple "reduction" loop is implemented.