previous_group previous up next next_group
Previous: 3. Overflow and Timeout Up: 7. Results for Convex Next: 5. Conclusion

4. Arithmetic Precision: 64-bit versus 32-bit

Since C3C64 is clearly the winner, in this section we only compare it with its 32-bit version, denoted C3C32. 6_fig:PerfectClub_convex_hull_dimension_C3C64_C3C32 illustrates these comparisons for PerfectClub convex hull database. C3C32 is faster than C3C64 for $37$ percent of all tests with three times faster for total run time of all tests, whereas C3C64 is faster for only $8$ percent of tests. SPEC95 convex hull database gives us similar results.

Figure 42: PerfectClub: Dimension C3C64-bit vs C3C32-bit in full database
\begin{figure}
\centering\epsfig {file=POLYBENCH_evaluations/PerfectClub_C3C64_C...
...05_DATABASE/dimension_C3C64_C3C32_crite.eps,height=5.9cm,width=14cm}\end{figure}


Table 23: SPEC95: Numbers of exceptions in full database
  SPEC95
timeout = 2 minutes #overflows #timeouts #operations
C3C64-bit 0 0 24151
C3C32-bit 32 0 24151


6_tab:SPEC95_convex_hull_64_32_exception shows that for the set of SPEC95 convex hull database, the $C^3$ partial factorization implementations, there are a few differences ($0$ to $32$ exceptions with $24151$ tests). For the PerfectClub, there are no differences with $19683$ tests. Our conclusion for this part is that 64-bit computation is preferable to 32-bit computation.


previous_group previous up next next_group
Previous: 3. Overflow and Timeout Up: 7. Results for Convex Next: 5. Conclusion
Nguyen Que Duong
2006-09-16