GMPbench result for GMP 6.0.0a on Core i7

[Andrew Mattingly]

The CPU was running in turbo mode at 3.9 GHz for the previously reported 
runs - my mistake.  The performance does drop for 4 parallel runs, but 
only by about 5%, because the effective turbo frequency drops to 3.7 GHz 
for 4 active cores (confirmed using "i7z" https://code.google.com/p/i7z/ . 
 When I disable turbo mode and run a single test at 3.4 GHz (technically, 
3391 MHz), I get a more modest result (see corrected table below):

    GMP     --------------- base -------------  ---- app -----    GMP 
Score/  CPUfreq  parallel
  version   multiply   divide      gcd  gcdext     rsa    pi     bench GHz 
   GHz      runs
 
    5.0.1    51903    48781     8575.1  5812.5  7068.3  49.563  3933.9 
1009     3.9       1
    6.0.0    65407    58715    10752    7354.2  8311.3  60.143  4789.4 
1228     3.9       1
    6.0.0    61389    55648    10218    6984.5  7886.8  56.854  4532.1 
1224     3.7       4

    5.0.1    45185    42562     7485.9  5064.6  6165.2  43.214  3430.0 
1009     3.4       1
    6.0.0    57008    51178     9374.9  6418.7  7251.0  52.478  4177.2 
1229     3.4       1

Regards,
Andrew.

>> Andrew Mattingly writes:
>>
>>  Comparing GMP 5.0.1 to GMP 6.0.0a using:
>> 
>>   benchmark: GMPbench 0.2 
>>   processor: Intel Core i7-4770K (Haswell 8MB L3) 3.4 GHz
>>   compiler:  "gcc 4.8.2" -O2 -m64 -mtune=corei7 -march=corei7
>> 
>>  produces the following results (best of 5 runs each):
>> 
>>    GMP     --------------- base -------------  ---- app -----    GMP 
Score/
>>  version   multiply   divide      gcd  gcdext     rsa    pi     bench 
GHz
>> 
>>    5.0.1    51903    48781     8575.1  5812.5  7068.3  49.563  3933.9 
1157
>>    6.0.0    65407    58715    10752    7354.2  8311.3  60.143  4789.4 
1409
>> 
>>  %improve      26.0     20.4     25.4    26.5    17.6  21.3      21.7 
21.7
 
> These numbers are very good, by far the highest GMPbench figures I've
> ever seen.
>
> The figures are actually so good that I suspect the CPU is being
> overclocked (using what the manufacturers call "turbo") during the test
> runs.  If you run 4 separate GMPbench runs in parallel, does the
> performance remain this high, or does it drop around 10 percent?
>
> Torbjörn