Page 163 - A Practical Guide from Design Planning to Manufacturing

136 Chapter Five

this would unnecessarily penalize short instructions, which also tend
to be common.
The most important rule in designing for high performance is to make
the common case fast. There is no need to waste large amounts of design
time and die area on functions that are seldom used. This idea is sometimes
formally referred to as Amdahl’s law, which estimates the overall
speedup obtained by decreasing the time spent on some fraction of a job. 1

Amdahl’s law:

Overall speedup = old time = 1 + fraction
new time 1 − fraction speedup
n
If the fraction of execution time reduced is very small, then even a very
large speedup for that piece will have very little impact on overall per-
formance. The concept of Amdahl’s law is used again and again in microar-
chitecture. Resources are often better spent on small improvements for
the most common operations instead of large speedups for rare events.
Many microarchitectural ideas slow down rare occurrences if that will
speed up the common case.
Frequency is the same for all instructions, so Amdahl’s law applies
only to improvements in IPC. More functional units allow higher issue
width. Better reordering algorithms reduce pipeline stalls. More spe-
cialized logic reduces the latency of computation. Larger caches can
reduce the average latency of memory accesses. All of these microarchi-
tectural changes improve IPC but at the cost of design complexity and
die area. Fred Pollack compared the relative performance of different
Intel processors and found that in the same fabrication technology, per-
formance improved roughly with the square root of added die area. This
relationship is often called Pollack’s rule. 2
Pollack’s rule:

Performance ∝ die area

This rule of thumb means that a new microarchitecture that requires
twice the die area will likely provide only a 40 percent improvement in
performance. The large amounts of new transistors needed for higher
performance microarchitectures have been made possible by Moore’s law.
Roadblocks to fabricating ever-smaller transistors may continue to be
overcome, but continued increases in microarchitectural complexity

1
Amdahl, “Validity of the single-processor approach.”
2
Pollack, “New Microarchitecture Challenges.”