HeavyHemi

kelkel1
Looking around the internet, I see others with 9900K scoring well over 500GFlops in Linpack at 5000MHz.
 
The best I have done is low 300GFlops at 5200 MHz.
 
Any advice?

Are you using the exact same linpack libraries they are and are you using an AVX offset in the BIOS? Both of those will significantly impact the score.

qswaaa
make Ram 4500 cl17

kelkel1

HeavyHemi

kelkel1
Looking around the internet, I see others with 9900K scoring well over 500GFlops in Linpack at 5000MHz.
 
The best I have done is low 300GFlops at 5200 MHz.
 
Any advice?

Are you using the exact same linpack libraries they are and are you using an AVX offset in the BIOS? Both of those will significantly impact the score.

Not sure about the binaries, but same linpack version.
 
My AVX offset is 0.

kelkel1

Kylearan

kelkel1

HeavyHemi

kelkel1
Looking around the internet, I see others with 9900K scoring well over 500GFlops in Linpack at 5000MHz.
 
The best I have done is low 300GFlops at 5200 MHz.
 
Any advice?

Are you using the exact same linpack libraries they are and are you using an AVX offset in the BIOS? Both of those will significantly impact the score.

Not sure about the binaries, but same linpack version.
 
My AVX offset is 0.

Please use LinX 0.9.5 or 0.9.6.  Not 0.6.5.
If you were the same person posting on OCN, that was 0.6.5 which doesn't even support AVX2, much less any modern processor correctly !

Using LinX 0.9.5 or Linpack Extreme 1.1.1.

kelkel1
 
 
Various VCore, including Auto.
16GB RAM.
In Linx, it was 40K something sample size.
Linpack Extreme 1.1.1 only gives ~170GFlops.
Delidded, direct die. At 5200 MHz, CPU Package hits 78°C with 28°C ambient.
Have not tried smaller sample size.

kelkel1
Just ran at 4500MHz and 4200MHz, as you suggested. 299 GFlops.

Kylearan

kelkel1
Thanks.
 
That VCore is crazy low.
 
I need 1.33 VCore to run LinX 0.9.5 at 5000MHz.

Remember he's using an AVX offset of -3.
LinX is running at 4.7 ghz on his system.  Then 5 ghz when its finished.

tekleadm
As others have mentioned, you have a throttling issue most likely, be it from power or thermal.
 
First, some background. The linpack benchmark is a good choice for stress testing, but most of these repackaged front-end tools either use outdated Intel Linpack executables, wrong configuration, or a poor implementation. On Intel cpu's, your best choice is to download Intel's linpack program directly from their site (https://software.intel.com/en-us/articles/intel-mkl-benchmarks-suite). These are continuously updated for each new generation of Intel cpus and highly tuned for them. They either perform poorly or not at all on AMD cpus. When you run it, the proper values for AVX testing are to set the "number of equations" and "leading dimension" to an odd multiple of 16 and have "leading dimension" >= "number of equations", that is, greater than or equal to. For 8th and 9th gen Intel, the number of vector units and width result in an ideal GFLOPs of "core speed" x 16 per core (not thread). For example, at 5 GHz, you would expect an ideal of 80 GFLOPs per core. Intel's linpack benchmark fully utilizes each core so hyperthreading may provide little or even negative benefit. If I remember correctly, by default, it will run threads up to the number of cores, and will not utilize hyperthreads.
 
Second, actual usage and testing. Intel linpack benchmark is a command line tool. It takes a few parameters as input while running it, but does not ask for the number of threads. Instead, you must set the number of threads using an environment variable. On Windows, just execute "set OMP_NUM_THREADS=X" where "X" is the number of threads, and then run the "linpack_xeon64.exe" (assuming you are on 64-bit). Here are my results on an i5-8600k at 5 GHz (0 AVX offset):
 
Single-threaded performance:
Average = 74.5295 (93% of ideal)
Size   LDA    Align. Time(s)    GFlops   Residual     Residual(norm) Check
20016  20016  4      71.746     74.5263  3.712881e-10 3.285072e-02   pass
20016  20016  4      71.752     74.5198  3.712881e-10 3.285072e-02   pass
20016  20016  4      71.730     74.5423  3.712881e-10 3.285072e-02   pass


Multi-threaded performance:
Average = 374.8176 (78% of ideal, ratio 5.03x ratio)
Size   LDA    Align. Time(s)    GFlops   Residual     Residual(norm) Check
20016  20016  4      14.273     374.6117 3.547637e-10 3.138867e-02   pass
20016  20016  4      14.268     374.7390 3.547637e-10 3.138867e-02   pass
20016  20016  4      14.255     375.1020 3.547637e-10 3.138867e-02   pass
 
Third, interpreting these results. To ensure stability, the check column must say pass, which signals that the error value is low. Another thing to check is that each result generated from each trial is identical, since it is repeating the exact same problem. Now, running the single-threaded test is just to set a baseline for performance. You should not be power-limited in this scenario, possibly thermal-limited depending on cooling. My 8600k was around 70c for these tests and performed near ideal. (You would see better performance in this case depending on cache overclock and RAM speed since the 20016 problem size uses about 3.2 GB of RAM.) Next, in the multi-threaded test, the performance increased, however relative to the single-threaded test, a lot of performance was lost in scaling. I only receive 78% of the ideal performance, and even in scaling from 1 to 6 threads, the performance only increased by a little over 5x. While there is some overhead in multithreading, this much lost means that I too am power-limited. My overclock is still stable however, and at 1.4V, I am reaching about 85c during that test so I have little room to increase it.
 
Now the 9900k has 8 cores, so assuming similar scaling in multithreading, 78% of (8 x 80 = 640 GFLOPS) is 500 GFLOPS. However, the 9900k has more cache so and usually is paired with higher speed RAM (versus mine which is 3000MHz - CL15), so a value over 500 GFLOPs is highly expected.
 
The last bit I will say is that tuning for Linpack performance means tuning for a full AVX heavy workload. Unless you are constantly rendering, encoding, or running scientific programs, you will be leaving performance that would be stable for gaming on the table since the AVX units are power-hungry and will limit your overclock ceiling. Using them will increase the power usage as well as the thermal load versus sticking to the scalar and SSE units.

Kylearan
What were your Gflops?
 
The one problem with the Z390 Dark is it doesn't support current monitoring.  If the VRM's supported monitoring, then HWinfo64 should also show the values.  So there isn't a clear way to determine how much current LinX is pulling.
At 35000 sample size and 5 ghz, you should be pulling between 160 to 200 amps, depending on vcore.  That would get you about 500 to 510 gflops, and good luck cooling that processor with that current load.
300 gflops is like you are not even using AVX.
I'm not familiar with the Dark Bios.  Weren't there some settings for current protection or VRM load capacity or something?