Superscalar out the Wazoo and other stories from Hot Chips VII

by Bradley C. Kuszmaul

At Hot Chips VII, held at Stanford on August 13-15 1995, everybody and his brother had a Tomasulo-style superscalar chip or something "just as good". `Everybody and his brother' included Intel (P6), HP (PA-8000), HaL (Sparc64, Sparc64+), Sun (UltraSparc), MIPS (R10000), AMD (29K, K5), and Cyrix (M5), but not IBM/Motorola PPC.

Intel chairman Gordon Moore talked about how expensive it is to build VLSI chips. The combination of high-performing x86 processors with high cost VLSI has lead many people to believe that Intel will take over the world.

Other important threads include: MPEG, Graphics and compression support on processor chips, low power support special-purpose chips, and embedded systems processors.

One of the most interesting results presented came from the SUIF compiler team [Monica Lam et al] at Stanford. They now claim to hold the record for "largest SpecFP92 number". The SUIF compiler can achieve 900 SpecFP92 running on an 8 processor DEC alpha workstation. (A single alpha processor achives about 400 SpecFP92). Their work shows that there is parallelism in ordinary programs that cannot be found by a single superscalar processor.

At this SPAM talk, I will describe what I heard and saw at Hot Chips VII, and I will conclude by describing what I think can be done to better exploit parallelism in ordinary C programs.