[Echidna]

Case V18sp-5
BGA
Transient, nonlinear, structural, 3 cumulative iteration
Solver "sparse, real-value, symmetric"
2073970 nodes ; 1266427 elements

in core solution


-------------------------------------------------------------------------


2x E5-2650V2 @ 3Ghz (16cores total) (For some reason my CPUs stopped at 3GHz max and didn't go to 3.2GHz)
128GB DDR3 RAM (not perfectly populated due to MB issues)

CPU Time 1 iteration(sec) = 270.6
ELAPSED Time 1 iteration(sec) = 271.3
ELAPSED Time 914.000

computational rate (gflops) for solve = 0.800521 // 0.798933
effective I/O rate (GB/sec) for solve = 2.982410 // 2.976493


---------------------------------------------------------------------------

DUKE711
Epyc 7551 @ 2,55
Memory total in MB-> 87124

CPU Time 1 iteration(sec) = 337.578
ELAPSED Time 1 iteration(sec) = 338.200
ELAPSED Time 1209.125

computational rate (mflops) for solve = 18414.8072 // 18384.2639
effective I/O rate (MB/sec) for solve = 70160.4145 // 70044.0444

-----------------------------------------------------------------------------------


DUKE711

Core-7 6850k @ 4,0:
Memory total in MB-> 73015

CPU Time 1 iteration(sec) = 276.281
ELAPSED Time 1 iteration(sec) = 277.488
ELAPSED Time 1011.047

computational rate (mflops) for solve = 8489.0170 // 8474.9668
effective I/O rate (MB/sec) for solve = 32343.1543 // 32289.6233

[flotus1]

2x E5-2650V2 @ 3Ghz (16cores total) (For some reason my CPUs stopped at 3GHz max and didn't go to 3.2GHz)

All-core turbo should be 3.0 GHz for these CPUs
https://www.microway.com/knowledge-c...dge-processor/

[Echidna]

Originally Posted by flotus1

All-core turbo should be 3.0 GHz for these CPUs
https://www.microway.com/knowledge-c...dge-processor/

Thanks! You are right.

[evcelica]

It's difficult to interpret these results without knowing core count? How many cores were all these benchmarks solved on? I'm mostly interested in the results for the EPYC processors.

Default 4 cores? Hopefully this is the case?
Or all cores on CPU? This is not a direct comparison, as with ANSYS you pay per core. so a core to core comparison would be much better.

Please let me know if anyone has any idea.
Here is the link to ANSYS's benchmark data for v18: https://www.ansys.com/fr-fr/solution...-benchmarks-18


After plotting all the data here, and comparing to the ANSYS mechanical benchmarks data, it looks like these benchmarks were done on all cores.

The XEON GOLD looks to be the best by far for these benchmarks. Though we don't have any data for EPYC ROME.
I can't compare results here with ANSYS data for sp-4 and 5. All ratings here are much higher than ANSYS, with no explanation.

Also, Looking at the top post on this page, how is the EPYC apparently slower than the 6850K, even though I/O and computational rate are over double the 6850K? And why is the reported I/O and solve rate so low for the E5 system, even though it has the lowest solve time?

[Duke711]

Originally Posted by evcelica

Also, Looking at the top post on this page, how is the EPYC apparently slower than the 6850K, even though I/O and computational rate are over double the 6850K? And why is the reported I/O and solve rate so low for the E5 system, even though it has the lowest solve time?

All run on max core count


Case V18cg-1
Static, linear, thermal
Solver "JCG, real value, symmetric"



Epyc 7551 @ 2,55

Multiply with A Memory Bandwidth : 101.03 GB/s
Multiply with A MFLOP Rate : 10848.33 MFlops
TOTAL PCG SOLVER SOLUTION CP TIME = 334.27 secs
TOTAL PCG SOLVER SOLUTION ELAPSED TIME = 334.49 secs

Core-7 6850k @ 4,0:

Multiply with A Memory Bandwidth : 57.44 GB/s
Multiply with A MFLOP Rate : 6167.20 MFlops
TOTAL PCG SOLVER SOLUTION CP TIME = 567.59 secs
TOTAL PCG SOLVER SOLUTION ELAPSED TIME = 568.02 secs


6850K @ GTX 780 6 GB --> with GPU accerlation!!!!!

Multiply with A Memory Bandwidth : 221.38 GB/s
Multiply with A MFLOP Rate : 23770.00 MFlops
TOTAL PCG SOLVER SOLUTION CP TIME = 232.23 secs
TOTAL PCG SOLVER SOLUTION ELAPSED TIME = 232.20 secs



--------------------------------------------------------


GTX 780 hardmodded to a quadro k5200 to be able activate the GPU accerlation.




-------------------


think the best Setup for



JCG (thermal) Solver is


- eight core cpu with a Tesla and lot of vram, in fea the graphic card run only in medium or low power mode





PCG Solver or iterative is


- big cpu with a lot of core's




Sparse Solver or direct is



- six core cpu with low power consumption




For fea is a multi cpu plattform with very high power consumption not a good idea. Only for CFD.