Instruction-Level Parallelism and Its Exploitation - Computer Architecture: A Quantitative Approach - page 241

We might expect that these two processors, implemented in the same technology and with
the same instruction set, would exhibit predictable behavior, in terms of relative performance
and energy consumption, meaning that power and performance would scale close to linearly.
We examine this hypothesis using three sets of benchmarks. The first sets is a group of Java,
single-threaded benchmarks that come from the DaCapo benchmarks, and the SPEC JVM98
benchmarks (see Esmaeilzadeh et al. [2011] for a discussion of the benchmarks and measure-
ments). The second and third sets of benchmarks are from SPEC CPU2006 and consist of the
integer and FP benchmarks, respectively.
As we can see in Figure 3.45 , the i7 significantly outperforms the Atom. All benchmarks are
at least four times faster on the i7, two SPECFP benchmarks are over ten times faster, and one
SPECINT benchmark runs over eight times faster! Since the ratio of clock rates of these two
processors is 1.6, most of the advantage comes from a much lower CPI for the i7: a factor of 2.8
for the Java benchmarks, a factor of 3.1 for the SPECINT benchmarks, and a factor of 4.3 for
the SPECFP benchmarks.