sr5sieve Benchmark thread

[michaf]

My Athlon 64 3800+ X2 processor gives the following results:

23000 p/sec on one processor
2*22700 p/sec on two processors

Cheers, Micha

[rogue]

With version 1.2.0, the speed is now at 60950 p/sec for 307 k's on my G5.

Unfortunately it is dedicated to other project at the moment, so I can't use it for sieving.

[michaf]

sr5sieve Version 1.2.0
AMD Athlon 64 X2 3800 (OC'd to 2200 MHz)
range 680-690
303 k's (7.06M candidates)

2*50800 p/sec.... Thats 101600 p/sec if I dedicate it to sieving alone!

[axn]

Does anyone have any benchmark for Pentium D 940 (3.40 GHz)?

[S485122]

D940 is 3.20 GHz
D945 and D950 are 3.40 GHz the first whithout virtualisation technology.

And indded on the Gimps Benchmark page you will only find the D945 (the line starting with Pentium 4 D 3400 2048) and the D930 (the line starting with Pentium 4 D 3000 2048).

But on the following page http://mersenne-aries.sili.net/bench.php you will find D9xx 3.20 GHz benchmarks (the lines starting with Intel(R) Pentium(R) D CPU 3.20GHz 3200 2048)

[axn]

Originally Posted by Jacob Visser

D940 is 3.20 GHz
D945 and D950 are 3.40 GHz the first whithout virtualisation technology.

And indded on the Gimps Benchmark page you will only find the D945 (the line starting with Pentium 4 D 3400 2048) and the D930 (the line starting with Pentium 4 D 3000 2048).

But on the following page http://mersenne-aries.sili.net/bench.php you will find D9xx 3.20 GHz benchmarks (the lines starting with Intel(R) Pentium(R) D CPU 3.20GHz 3200 2048)

Yes, I meant 945. Actually, I need data for sr5sieve 1.3.x version

[geoff]

Here are some benchmarks for sr2sieve 1.4.38 (linux-x86 binary). With the improvements to the 32-bit SSE2 code the hyperthreaded P4 now matches the P3 clock for clock with SoB.dat and Riesel.dat. I think the reason it is still relatively slower with sr5data.txt is that the SSE2 code is concentrated in the baby-steps giant-steps routine, but proportionally less time is spent there when the range of n is small.

M1 = Pentium 3 @ 600MHz (Coppermine EB, 16Kb L1, 256Kb L2).
M2 = Pentium 4 @ 2.9GHz (Northwood C, 8Kb L1, 512Kb L2), single thread.
M3 = Pentium 4 @ 2.9GHz (Northwood C, 8Kb L1, 512Kb L2), two hyperthreads.

                         19k SoB.dat    68k riesel.dat    237k sr5data.txt
                         -----------    --------------    ----------------
proth_sieve_cmov 0.42
 M1 p=100e12              86 kp/s        31 kp/s

proth_sieve_sse2 0.42
 M2 p=100e12             201 kp/s        83 kp/s
 M3 p=100e12             330 kp/s       133 kp/s

sr2sieve-intel 1.4.38
 M1 p=100e12              92 kp/s        53 kp/s           30 kp/s
 M2 p=100e12             293 kp/s       161 kp/s           77 kp/s
 M3 p=100e12             451 kp/s       244 kp/s          113 kp/s

Some rough timings on a Celeron D 2.8GHz running Windows say that this version is about 10% faster than JJsieveSSE2 on the 19k SoB.dat, but the times vary a lot from run to run.

Any times for the 32-bit binary on Core2 or Athlon64 machines would be welcome. I don't yet know how well the SSE2 code suits these machines.

[geoff]

I finally gave in and bought a new machine, so here are some benchmarks for a 2.67 GHz Core 2 Duo (E6750, 32Kb L1, 4Mb L2):

                                19k SoB.dat    68k riesel.dat    237k sr5data.txt
                                -----------    --------------    ----------------
1.5.15 64-bit p=100e12          1115 kp/s       606 kp/s          300 kp/s
1.5.15 32-bit p=100e12           729 kp/s       384 kp/s          184 kp/s

[geoff]

The SoB.dat and Riesel.dat benchmarks above were accidentally done with p=1000e12 instead of 100e12. Here are the results with p=100e12:

                            19k SoB.dat    68k riesel.dat    237k sr5data.txt
                            -----------    --------------    ----------------
1.5.15 64-bit p=100e12      1160 kp/s       626 kp/s          300 kp/s
1.5.15 32-bit p=100e12       689 kp/s       368 kp/s          184 kp/s

The 64-bit version becomes slower as p increases, while the 32-bit version becomes faster. I think this is due to the lower precision of the SSE2 floating point operations used in the 64-bit version leading to more rounding corrections being needed as p increases.

[Xentar]

I just tested the sr5sieve AMD version and intel version, at the same time.

System:
Windows XP Prof 32 bit; Intel C2D E6600 @2.4 Ghz (original speed - not overclocked)

using original sr5.dat file (232k; size 19.7MB), sr5sieve 1.5.17

148 - 150 kp/sec with -Intel.exe
146 - 147 kp/sec with -AMD.exe
then I closed the -AMD software, and continued with -Intel: 149 kp/sec

so, Intel version is a little bit faster.

[rogue]

12200-12300 done. It took about two days using both cores on a Core 2 Duo