Unusually low P4 equiv for EPYC in VM?

[haruka]

I'm utilizing some free computational resource of a server with EPYC 7313, and the CPU is shown as:

Model AMD EPYC 7313 16-Core Processor
Features 16-core, hyperthreaded, Prefetch,SSE,SSE2,SSE4,AVX,AVX2,FMA,
Speed 2.999 GHz (2.775 GHz P4 effective equivalent)
L1 Cache 64 KB
L2 Cache 512 KB
L3 Cache 16,384 KB

on the CPU page.

The cache amount is incorrect because this is running in a VM, but I've optimized the settings so there is very little performance penalty for that.

However, the P4 equivalent seems completely wrong with only 2.775 GHz, which hints that one CPU core is even slower than a P4. (I know "Hours per day" affects this, but it's currently set as 24 hours.)

I suppose this would negatively affect what kind of work my machine will receive automatically and I would need to manually select PRP work if I want to run it?

How is the equivalent calculated anyway? This figure just feels way too off, and I'd like to know more about the logic behind that.

[Mark Rose]

You'll get better performance if a worker uses only a single chiplet.

That number is very low. Was the server busy at the time?

Note that even though mprime runs niced, its memory bandwidth consumption can create massive latency for other things running on the same CPU.

[Prime95]

This AMD CPU has been identified as Bulldozer architecture. Bulldozer was horrible.
If this CPU is actually based on Zen architecture, I'll need to know the family code and model code as returned by CPUID. Not sure which program might easily output this. Prime95 has a hidden code to do that -- I can look that up if you need it.

[haruka]

Originally Posted by Mark Rose

Note that even though mprime runs niced, its memory bandwidth consumption can create massive latency for other things running on the same CPU.

That's a great point, I'll need to do more testing on that later.

Originally Posted by Prime95

This AMD CPU has been identified as Bulldozer architecture. Bulldozer was horrible.
If this CPU is actually based on Zen architecture, I'll need to know the family code and model code as returned by CPUID. Not sure which program might easily output this. Prime95 has a hidden code to do that -- I can look that up if you need it.

I decided to write a program to output the raw values:

#include <stdio.h>

int main() {
    volatile int a = 1;
    __asm__ (
        ".intel_syntax noprefix \n"
        "mov eax, 1 \n"
        "cpuid \n"
        "mov %0, eax \n"
        ".att_syntax \n"
        : "=r" (a)
        :
        : "eax"
    );
    int family = (a & (0b1111 << 8)) >> 8;
    int model = (a & (0b1111 << 4)) >> 4;
    int exfamily = (a & (0b11111111 << 20)) >> 20;
    int exmodel = (a & (0b1111 << 16)) >> 16;
    printf("f = %d\nxf = %d\nm = %d\nxm = %d\n", family, exfamily, model, exmodel);
    return 0;
}

which outputs:

f = 15
xf = 10
m = 1
xm = 0

However, it seems p95 should think this is Zen:

		else if (family_code == 15 && (extended_family == 8 || extended_family == 10))
			CPU_ARCHITECTURE = CPU_ARCHITECTURE_AMD_ZEN;

[Prime95]

My bad, the P4 equivalent number is determined by the server. The server has no access to the CPUID information. All the server has to go on is: "Model AMD EPYC 7313 16-Core Processor". I'll see if I can change the server to assume EPYC's are zen based.

I don't think your P4 equiv value will affect the assignments you get but I could be wrong (again).

[slandrum]

Originally Posted by Prime95

I don't think your P4 equiv value will affect the assignments you get but I could be wrong (again).

According to the assignment rules page https://www.mersenne.org/thresholds/ it can limit what assignments you can get.

[haruka]

Looks good now:

Speed 2.999 GHz (11.000 GHz P4 effective equivalent)