An In-Depth Look at Computer Performance Growth
Authors: Magnus Ekman, Fredrik Warg and Jim Nilsson
Abstract:

It is a common belief that computer performance growth is over 50% annually, or that performance doubles every 18-20 months. By analyzing publicly available results from the SPEC integer (CINT) benchmark suites, we conclude that this was true between 1985 and 1996 the early years of the RISC paradigm.

During the last 7.5 years (1996-2004), however, performance growth has slowed down to 41%, with signs of a continuing decline. Meanwhile, clock frequency has improved with about 29% annually. The improvement in clock frequency was enabled both by an annual device speed scaling of 20% as well as by longer pipelines with a lower gate-depth in each stage. This paper takes a fresh look at and tries to remove the confusion about performance scaling that exists in the computer architecture community.

Keywords: Performance evaluation.
Fulltext: pdf
Published: Technical report 04-9, Department of Computer Engineering, Chalmers University of Technology. Also published in ACM SIGARCH Computer Architecture News, Volume 33, Issue 1 (March 2005), pp. 144 - 147, 2005.

Update 2006-07-08: The performance growth graph shown below corresponds to Figure 2 in this paper, but is updated in April 2006. The assertion that single-thread performance growth will continue to slow down still holds. Between 2001 and 2006 the average annual performance growth is not higher than 29%.